pmacct [IP traffic accounting : BGP : BMP : RPKI : BGP-LS : Streaming Telemetry] pmacct is Copyright (C) 2003-2025 by Paolo Lucente TABLE OF CONTENTS: I. Daemons and plugins included with pmacct distribution II. Configuring pmacct for compilation and installing III. Brief SQL (MySQL, PostgreSQL, SQLite 3.x) setup examples IV. Running the libpcap-based daemon (pmacctd) V. Running the NetFlow/IPFIX and sFlow daemons (nfacctd/sfacctd) VI. Running the NFLOG-based daemon (uacctd) VII. Running the pmacct IMT client (pmacct) VIII. Running the RabbitMQ/AMQP plugin IX. Running the Kafka plugin X. Internal buffering and queueing XI. Quickstart guide to packet/flow classification XII. Quickstart guide to setup a NetFlow/IPFIX agent/probe XIII. Quickstart guide to setup a sFlow agent/probe XIV. Quickstart guide to setup the BGP daemon XV. Quickstart guide to setup a NetFlow/IPFIX/sFlow replicator XVI. Quickstart guide to setup the BMP daemon XVII. Quickstart guide to setup Streaming Telemetry collection XVIII. Running the print plugin to write to flat-files XIX. Quickstart guide to setup GeoIP lookups XX. Using pmacct as traffic/event logger XXI. Connecting pmacct to a Redis cache XXII. Miscellaneous notes and troubleshooting tips I. Daemons and plugins included with pmacct distribution All traffic accounting daemons can print statistics to stdout, keep them in memory tables, store persistently to open-source RDBMS (MySQL, PostgreSQL, Sqlite 3) or to noSQL databases (ie. BerkeleyDB) and to flat-files, and publish to AMQP and Kafka brokers (typically to insert in ElasticSearch, InfluxDB, Druid, ClickHouse and, more in general, all backends which are not natively supported by pmacct). BGP, BMP and Streaming Telemetry daemons can publish control and infrastructure planes to AMQP and Kafka brokers. This is a list of the daemons included in the pmacct distribution: pmacctd libpcap-based accounting daemon: it captures packets from one or multiple interfaces it is bound to. Other than acting as a collector, this daemon can also export statistics via NetFlow, IPFIX and sFlow protocols. nfacctd NetFlow/IPFIX accounting daemon: it listens for NetFlow v5/v9 and IPFIX packets on one or more interfaces (IPv4 and IPv6). Other than acting as a collector, this daemon can also replicate to 3rd party collectors. sfacctd sFlow accounting daemon; it listens for sFlow packets v2, v4 and v5 on one or more interfaces (both IPv4 and IPv6). Other than acting as a collector, this daemon can also replicate to 3rd party collectors. uacctd Linux Netlink NFLOG accounting daemon; it captures packets by leveraging a NFLOG multicast group - and works only on Linux. Other than acting as a collector, this daemon can also export statistics via NetFlow, IPFIX and sFlow protocols. pmtelemetryd Standalone Streaming Telemetry collector daemon; listens for telemetry data binding to a TCP or UDP port and logs real-time and/or dumps at regular time-intervals to configured backends. pmbgpd Standalone BGP collector daemon; acts as a passive iBGP or eBGP neighbor and maintains per-peer RIBs; can log real-time and/or dump at regular time-intervals BGP data to configured backends. pmbmpd Standalone BMP collector daemon; can log real-time and/or dump at regular time-intervals BMP/BGP data to configured backends. pmacct commandline pmacct client; it allows to retrieve data from a memory table plugin; it can perform queries over data or do bulk data retrieval. Output is formatted, CSV or JSON format. suitable for data injection in 3rd party tools like RRDtool, Gnuplot or SNMP server among the others. Given its open and pluggable architecture, pmacct is easily extensible with new plugins. Here is a list of traffic accounting plugins included in the official pmacct distribution: memory data is stored in a memory table and can be fetched via the pmacct command-line client tool, 'pmacct'. This plugin also implements a push model and allows easily to inject data into 3rd party tools. The plugin is recommended for prototyping and proof-of-concept (without mass traffic generation) and smaller-scale / home production environments and is compiled in by default. mysql a working MySQL/MariaDB installation can be used for data storage. This plugin can be compiled using the --enable-mysql switch. pgsql a working PostgreSQL installation can be used for data storage. This plugin can be compiled using the --enable-pgsql switch. sqlite3 a working SQLite 3.x or BerkeleyDB 5.x (compiled in with the SQLite API) installation can be used for data storage. This plugin can be compiled using the --enable-sqlite3 switch. print data is printed at regular intervals to flat-files or standard output in tab-spaced, CSV and JSON formats. This plugin is compiled in by default. amqp data is sent to a RabbitMQ broker, running AMQP protocol, for delivery to consumer applications or tools. Popular consumers are ElasticSearch, InfluxDB, Druid and ClickHouse. This plugin can be compiled using the --enable-rabbitmq switch. kafka data is sent to a Kafka broker for delivery to consumer applications or tools. Popular consumers are ElasticSearch, InfluxDB, Druid and ClickHouse. This plugin can be compiled using the --enable-kafka switch. tee applies to nfacctd and sfacctd daemons only. It's a featureful packet replicator for NetFlow/IPFIX/sFlow data. This plugin is compiled in by default. nfprobe applies to pmacctd and uacctd daemons only. Exports collected data via NetFlow v5/v9 or IPFIX. This plugin is compiled in by default. sfprobe applies to pmacctd and uacctd daemons only. Exports collected data via sFlow v5. This plugin is compiled in by default. II. Configuring pmacct for compilation and installing The simplest way to configure the package for compilation is to download the latest stable released tarball from http://www.pmacct.net/ and let the configure script to probe default headers and libraries for you. The only dependency that pmacct brings is libpcap library and headers: libpcap-dev on Debian/Ubuntu, libpcap-devel on CentOS/RHEL (note: this may need enabling extra yum repos!) or (self-compiled) equivalent must be installed on the system. Then, a first round of guessing is done via pkg-config then, for some libraries, "typical" default locations are checked, ie. /usr/local/lib. Switches one likely wants enabled are already set so, ie. 64 bits counters and multi-threading (pre- requisite for the BGP and BMP daemon codes); the full list of switches enabled by default are marked as 'default: yes' in the "./configure --help" output. SQL plugins, AMQP and Kafka support are all disabled by default instead. A few examples will follow; to get the list of available switches, you can use the following command-line: shell> ./configure --help Examples on how to enable support for (1) MySQL, (2) PostgreSQL and (3) SQLite: (1) libmysqlclient-dev package or (self-compiled) equivalent being installed: shell> ./configure --enable-mysql (2) libpq-dev package or (self-compiled) equivalent being installed: shell> ./configure --enable-pgsql (3) libsqlite3-dev package or (self-compiled) equivalent being installed: shell> ./configure --enable-sqlite3 If cloning the GitHub repository ( https://github.com/pmacct/pmacct ) instead, the configure script has to be generated, resulting in one extra step than the process just described. Please refer to the Building section of the README.md document for instruction about cloning the repo, generate the configure script along with the required installed packages. Then compile and install simply typing: shell> make; make install Should you want, for example, to compile pmacct with PostgreSQL support and have installed PostgreSQL in /usr/local/postgresql and pkg-config is unable to help, you can supply this non-default location as follows (assuming you are running the bash shell): shell> export PGSQL_LIBS="-L/usr/local/postgresql/lib -lpq" shell> export PGSQL_CFLAGS="-I/usr/local/postgresql/include" shell> ./configure --enable-pgsql If the library does actually support pkg-config but the .pc pkg-config file is in some non-standard location, this can be supplied as follows: shell> export PKG_CONFIG_PATH=/usr/local/postgresql/pkgconfig/ shell> ./configure --enable-pgsql Special case is to compile pmacct with MySQL support but MySQL is installed in some non-default location. MySQL brings the mysql_config tool that works similarly to pkg-config. Make sure the tool is on the path so that it can be executed by the configure script, ie.: shell> export PATH=$PATH:/usr/local/mysql/bin shell> ./configure --enable-mysql By default all tools - flow, BGP, BMP and Streaming Telemetry - are compiled. Specific tool sets can be disabled. For example, to compile only flow tools (ie. no pmbgpd, pmbmpd, pmtelemetryd) the following command-line can be used: shell> ./configure --disable-bgp-bins --disable-bmp-bins --disable-st-bins Once daemons are installed you can check: * Basic instrumenting of each daemon via its help page, ie.: shell> pmacctd -h * Review daemon version and build details, ie.: shell> sfacctd -V * Check supported traffic aggregation primitives and their description, ie.: shell> nfacctd -a IIa. Compiling pmacct with JSON support JSON encoding is supported via the Jansson library (http://www.digip.org/jansson/ and https://github.com/akheron/jansson); a library version >= 2.5 is required. To compile pmacct with JSON support simply do: shell> ./configure --enable-jansson However should you have installed Jansson in the /usr/local/jansson directory and pkg-config is unable to help, you can supply this non-default location as follows (assuming you are running the bash shell): shell> export JANSSON_LIBS="-L/usr/local/jansson/lib -ljansson" shell> export JANSSON_CFLAGS="-I/usr/local/jansson/include" shell> ./configure --enable-jansson IIb. Compiling pmacct with Apache Avro support Apache Avro encoding is supported via libavro library (http://avro.apache.org/ and https://avro.apache.org/docs/1.9.1/api/c/index.html); Avro depends on the Jansson JSON parser version 2.3 or higher so please review the previous section "Compiling pmacct with JSON support"; then, to compile pmacct with Apache Avro support simply do: shell> ./configure --enable-jansson --enable-avro However should you have installed libavro in the /usr/local/avro directory and pkg-config is unable to help, you can supply this non-default location as follows (assuming you are running the bash shell): shell> export AVRO_LIBS="-L/usr/local/avro/lib -lavro" shell> export AVRO_CFLAGS="-I/usr/local/avro/include" shell> ./configure --enable-kafka --enable-jansson --enable-avro IIc. Compiling pmacct against a own libpcap library Compiling against a downloaded libpcap library may be wanted for several reasons including the version packaged with the Operating System is too old, a custom libpcap library needs to be compiled (ie. with support for PF_RING) or static linking is wanted. Once libpcap is downloaded, if static linking is wanted (ideal for example for distributing pmacct without external dependencies), the library can be configured for compiling: shell> ./configure --disable-so Which passes the compiler the '-static' knob. pmacct should be pointed to the own libpcap library when configuring for compiling: shell> ./configure --with-pcap-libs=/path/to/libpcap-x.y.z --with-pcap-includes=/path/to/libpcap-x.y.z Once pmacct is compiled, it can be confirmed that the right library was picked by doing, for example, a 'pmacctd -V' and seeing the version of libpcap matches with the supplied version. It has to be noted however that static compiling on GNU systems can yeld variable success; the recommendation is to consider containers first (look into the 'docker/' directory). A use-case for a PF_RING-enabled libpcap is that by hashing and balancing collected traffic over multiple NIC queues (ie. if using Intel X520) it is possible to scale pmacctd horizontally, with one pmacctd instance reading from one or multiple queues. The queues can be managed via the 'ethtool' tool (ie. 'ethtool -l enp1s0f0' to list, 'ethtool -L enp1s0f0 combined 16' to access 16 queues, etc.) and pmacctd can be bound to a single queue, ie. 'pmacctd -i enp1s0f0@0', or multiple ones via a pcap_interfaces_map, ie. ifname=enp1s0f0@0 ifindex=100 ifname=enp1s0f0@1 ifindex=101 ifname=enp1s0f0@2 ifindex=102 ifname=enp1s0f0@3 ifindex=103 III. Brief SQL and noSQL setup examples RDBMS require a table schema to store data. pmacct offers two options: use one of the few pre-determined table schemas available (sections IIIa, b and c) or compose a custom schema to fit your needs (section IIId). If you are blind to SQL the former approach is recommended, although it can pose scalability issues in larger deployments; if you know some SQL the latter is definitely the way to go. Scripts for setting up RDBMS are located in the 'sql/' tree of the pmacct distribution tarball. For further guidance read the relevant README files in such directory. One of the crucial concepts to deal with, when using default table schemas, is table versioning: please read more about this topic in the FAQS document (Q17). IIIa. MySQL examples shell> cd sql/ - To create v1 tables: shell> mysql -u root -p < pmacct-create-db_v1.mysql shell> mysql -u root -p < pmacct-grant-db.mysql Data will be available in 'acct' table of 'pmacct' DB. - To create v2 tables: shell> mysql -u root -p < pmacct-create-db_v2.mysql shell> mysql -u root -p < pmacct-grant-db.mysql Data will be available in 'acct_v2' table of 'pmacct' DB. ... And so on for the newer versions. IIIb. PostgreSQL examples Which user has to execute the following two scripts and how to authenticate with the PostgreSQL server depends upon your current configuration. Keep in mind that both scripts need postgres superuser permissions to execute some commands successfully: shell> cp -p *.pgsql /tmp shell> su - postgres To create v1 tables: shell> psql -d template1 -f /tmp/pmacct-create-db.pgsql shell> psql -d pmacct -f /tmp/pmacct-create-table_v1.pgsql To create v2 tables: shell> psql -d template1 -f /tmp/pmacct-create-db.pgsql shell> psql -d pmacct -f /tmp/pmacct-create-table_v2.pgsql ... And so on for the newer versions. A few tables will be created into 'pmacct' DB. 'acct' ('acct_v2' or 'acct_v3') table is the default table where data will be written when in 'typed' mode (see 'sql_data' option in CONFIG-KEYS document; default value is 'typed'); 'acct_uni' ('acct_uni_v2' or 'acct_uni_v3') is the default table where data will be written when in 'unified' mode. Since v6, PostgreSQL tables are greatly simplified: unified mode is no longer supported and an unique table ('acct_v6', for example) is created instead. IIIc. SQLite examples shell> cd sql/ - To create v1 tables: shell> sqlite3 /tmp/pmacct.db < pmacct-create-table.sqlite3 Data will be available in 'acct' table of '/tmp/pmacct.db' DB. Of course, you can change the database filename basing on your preferences. - To create v2 tables: shell> sqlite3 /tmp/pmacct.db < pmacct-create-table_v2.sqlite3 Data will be available in 'acct_v2' table of '/tmp/pmacct.db' DB. ... And so on for the newer versions. IIId. Custom SQL tables Custom tables can be built by creating your own SQL schema and indexes. This allows to mix-and-match the primitives relevant to your accounting scenario. To flag intention to build a custom table the sql_optimize_clauses directive must be set to true, ie.: sql_optimize_clauses: true sql_table: aggregate: How to build the custom schema? Let's say the aggregation method of choice (aggregate directive) is "vlan, in_iface, out_iface, etype" the table name is "acct" and the database of choice is MySQL. The SQL schema is composed of four main parts, explained below: 1) A fixed skeleton needed by pmacct logics: CREATE TABLE ( packets INT UNSIGNED NOT NULL, bytes BIGINT UNSIGNED NOT NULL, stamp_inserted DATETIME NOT NULL, stamp_updated DATETIME ); 2) Indexing: primary key (of your choice, this is only an example) plus any additional index you may find relevant. 3) Primitives enabled in pmacct, in this specific example the ones below; should one need more/others, these can be looked up in the sql/README.mysql file in the section named "Aggregation primitives to SQL schema mapping:" : vlan INT(2) UNSIGNED NOT NULL, iface_in INT(4) UNSIGNED NOT NULL, iface_out INT(4) UNSIGNED NOT NULL, etype INT(2) UNSIGNED NOT NULL, 4) Any additional fields, ignored by pmacct, that can be of use, these can be for lookup purposes, auto-increment, etc. and can be of course also part of the indexing you might choose. Putting the pieces together, the resulting SQL schema is below along with the required statements to create the database: DROP DATABASE IF EXISTS pmacct; CREATE DATABASE pmacct; USE pmacct; DROP TABLE IF EXISTS acct; CREATE TABLE acct ( vlan INT(2) UNSIGNED NOT NULL, iface_in INT(4) UNSIGNED NOT NULL, iface_out INT(4) UNSIGNED NOT NULL, etype INT(2) UNSIGNED NOT NULL, packets INT UNSIGNED NOT NULL, bytes BIGINT UNSIGNED NOT NULL, stamp_inserted DATETIME NOT NULL, stamp_updated DATETIME, PRIMARY KEY (vlan, iface_in, iface_out, etype, stamp_inserted) ); To grant default pmacct user permission to write into the database look at the file sql/pmacct-grant-db.mysql IIIe. Historical accounting Enabling historical accounting allows to aggregate data in time-bins (ie. 5 mins, hour, day, etc.) in a flexible and fully configurable way. Two timestamps are available: the 'stamp_inserted' field, representing the basetime of the time-bin, and 'stamp_updated', the last time the time-bin was updated. Following a pretty standard config fragment to slice data into nicely aligned (or rounded-off) 5 minutes time-bins: sql_history: 5m sql_history_roundoff: m IIIf. INSERTs-only UPDATE queries are expensive; this is why, even if they are supported by pmacct, a savy approach would be to cache data for longer times in memory and write them off once per time-bin (sql_history): this results into a much lighter INSERTs-only setup. This is an example based on 5 minutes time-bins: sql_refresh_time: 300 sql_history: 5m sql_history_roundoff: m sql_dont_try_update: true Note that sql_refresh_time is always expressed in seconds. An alternative approach for cases where sql_refresh_time must be kept shorter than sql_history (for example because a) of long sql_history periods, ie. hours or days, and/or because b) near real-time data feed is a requirement) is to set up a synthetic auto-increment 'id' field: it successfully prevents duplicates but comes at the expenses of GROUP BYs when querying data. IV. Running the libpcap-based daemon (pmacctd) All deamons including pmacctd can be run with commandline options, using a config file or a mix of the two. Sample configuration files are in examples/ tree. Note also that most of the new features are available only as config directives. To be aware of the existing configuration directives, please read the CONFIG-KEYS document. Show all available pmacctd commandline switches: shell> pmacctd -h Run pmacctd reading configuration from a specified file (see examples/ tree for a brief list of some commonly useed keys; divert your eyes to CONFIG-KEYS for the full list). This example applies to all daemons: shell> pmacctd -f pmacctd.conf Daemonize the process; listen on eth0; aggregate data by src_host/dst_host; write to a MySQL server; filter in only traffic with source prefix 10.0.0.0/16; note that filters work the same as tcpdump so you can refer to libpcap/tcpdump man pages for examples and further reading about the supported filtering syntax. shell> pmacctd -D -c src_host,dst_host -i eth0 -P mysql src net 10.0.0.0/16 Or written the configuration way: ! daemonize: true plugins: mysql aggregate: src_host, dst_host pcap_interface: eth0 pcap_filter: src net 10.0.0.0/16 ! ... Print collected traffic data aggregated by src_host/dst_host over the screen; refresh data every 30 seconds and listen on eth0. shell> pmacctd -P print -r 30 -i eth0 -c src_host,dst_host Or written the configuration way: ! plugins: print print_refresh_time: 30 aggregate: src_host, dst_host pcap_interface: eth0 ! ... Print collected traffic data aggregated by src_host/dst_host over the screen; refresh data every 30 seconds and listen on eth0 and eth1, listed in the file pointed by pcap_interfaces_map (see 'examples/pcap_interfaces.map.example' for more advanced uses of the map): ! plugins: print print_refresh_time: 30 aggregate: src_host, dst_host pcap_interfaces_map: /path/to/pcap_interfaces.map pcap_ifindex: map ! ... Then in /path/to/pcap_interfaces.map: ! ifindex=100 ifname=eth0 ifindex=200 ifname=eth1 ! ... Daemonize the process; let pmacct aggregate traffic in order to show in vs out traffic for network 192.168.0.0/16; send data to a PostgreSQL server. This configuration is not possible via commandline switches; the corresponding configuration follows: ! daemonize: true plugins: pgsql[in], pgsql[out] aggregate[in]: dst_host aggregate[out]: src_host aggregate_filter[in]: dst net 192.168.0.0/16 aggregate_filter[out]: src net 192.168.0.0/16 sql_table[in]: acct_in sql_table[out]: acct_out ! ... And now enabling historical accounting. Split traffic by hour and write to the database every 60 seconds: ! daemonize: true plugins: pgsql[in], pgsql[out] aggregate[in]: dst_host aggregate[out]: src_host aggregate_filter[in]: dst net 192.168.0.0/16 aggregate_filter[out]: src net 192.168.0.0/16 sql_table[in]: acct_in sql_table[out]: acct_out sql_refresh_time: 60 sql_history: 1h sql_history_roundoff: h ! ... Let's now translate the same example in the memory plugin world. One of the use-cases for this plugin is when feeding 3rd party tools with bytes/ packets/flows counters. Examples how to query the memory table with the 'pmacct' client tool will follow later in this document. Now, note that each memory table need its own pipe file in order to get queried by the client: ! daemonize: true plugins: memory[in], memory[out] aggregate[in]: dst_host aggregate[out]: src_host aggregate_filter[in]: dst net 192.168.0.0/16 aggregate_filter[out]: src net 192.168.0.0/16 imt_path[in]: /tmp/pmacct_in.pipe imt_path[out]: /tmp/pmacct_out.pipe ! ... As a further note, check CONFIG-KEYS document about more imt_* directives as they will support in the task of fine tuning the size and boundaries of memory tables, if default values are not ok for your setup. Now, fire multiple instances of pmacctd, each on a different interface; again, because each instance will have its own memory table, it will require its own pipe file for client queries aswell (as explained in the previous examples): shell> pmacctd -D -i eth0 -m 8 -s 65535 -p /tmp/pipe.eth0 shell> pmacctd -D -i ppp0 -m 0 -s 32768 -p /tmp/pipe.ppp0 Run pmacctd logging what happens to syslog and using "local2" facility: shell> pmacctd -c src_host,dst_host -S local2 NOTE: superuser privileges are needed to execute pmacctd correctly. V. Running the NetFlow/IPFIX and sFlow daemons (nfacctd/sfacctd) All examples about pmacctd are also valid for nfacctd and sfacctd with the exception of directives that apply exclusively to libpcap. If you have skipped examples in the previous section, please read them before continuing. All config keys available are in the CONFIG-KEYS document. Some examples: Run nfacctd reading configuration from a specified file: shell> nfacctd -f nfacctd.conf Daemonize the process; aggregate data by sum_host (by host, summing inbound + outbound traffic); write to a local MySQL server. Listen on port 5678 for incoming Netflow datagrams (from one or multiple NetFlow agents): shell> nfacctd -D -c sum_host -P mysql -l 5678 Let's now configure pmacct to insert data in MySQL every two minutes, enable historical accounting with 10 minutes time-bins and make use of a SQL table version 4: ! daemonize: true plugins: mysql aggregate: sum_host nfacctd_port: 5678 sql_refresh_time: 120 sql_history: 10m sql_history_roundoff: mh sql_table_version: 4 ! ... Va. NetFlow daemon & accounting NetFlow v9/IPFIX options NetFlow v9/IPFIX can send option records other than flow ones, typically used to send to a collector mappings among interface SNMP ifIndexes to interface names or VRF ID's to VRF names or extra sampling information. nfacctd_account_options enables accounting of option records then these should be split from regular flow records. Below is a sample config: nfacctd_time_new: true nfacctd_account_options: true ! plugins: print[data], print[option_vrf], print[option_if], print[option_sampling] ! pre_tag_filter[data]: 100 aggregate[data]: peer_src_ip, in_iface, out_iface, tos, vrf_id_ingress, vrf_id_egress print_refresh_time[data]: 300 print_history[data]: 300 print_history_roundoff[data]: m print_output_file_append[data]: true print_output_file[data]: /path/to/flow_%s print_output[data]: csv ! pre_tag_filter[option_vrf]: 200 aggregate[option_vrf]: peer_src_ip, vrf_id_ingress, vrf_name print_refresh_time[option_vrf]: 300 print_history[option_vrf]: 300 print_history_roundoff[option_vrf]: m print_output_file_append[option_vrf]: true print_output_file[option_vrf]: /path/to/option_vrf_%s print_output[option_vrf]: event_csv ! pre_tag_filter[option_if]: 200 aggregate[option_if]: peer_src_ip, in_iface, int_descr print_refresh_time[option_if]: 300 print_history[option_if]: 300 print_history_roundoff[option_if]: m print_output_file_append[option_if]: true print_output_file[option_if]: /path/to/option_if_%s print_output[option_if]: event_csv ! pre_tag_filter[option_sampling]: 200 aggregate[option_sampling]: peer_src_ip, sampler_id, sampler_interval print_refresh_time[option_sampling]: 300 print_history[option_sampling]: 300 print_history_roundoff[option_sampling]: m print_output_file_append[option_sampling]: true print_output_file[option_sampling]: /path/to/option_sampling_%s print_output[option_sampling]: event_csv ! aggregate_primitives: /path/to/primitives.lst pre_tag_map: /path/to/pretag.map maps_refresh: true Below is the referenced pretag.map: set_tag=100 ip=0.0.0.0/0 sample_type=flow set_tag=200 ip=0.0.0.0/0 sample_type=option Below is the referenced primitives.lst: name=vrf_id_ingress field_type=234 len=4 semantics=u_int name=vrf_id_egress field_type=235 len=4 semantics=u_int name=vrf_name field_type=236 len=32 semantics=str ! name=int_descr field_type=83 len=64 semantics=str ! name=sampler_interval field_type=50 len=4 semantics=u_int name=sampler_id field_type=48 len=2 semantics=u_int Vb. Distributing NetFlow v9/IPFIX templates when clustering One of the possible ways to cluster multiple nfacctd daemons running on the same system (ie. to harness all [configured] CPU threads) is to rely on the Linux/BSD SO_REUSEPORT feature (read more in the "Miscellaneous notes and troubleshooting tips" section of this document). NetFlow v9/IPFIX protocols are template-based and templates are emitted in specific Options packets and are indeed needed in order to properly decode actual Data packets. Hence a load-balancing cluster (when not hashing but performing some sort of round-robin) does pose the problem of disseminating the templates across the clustered daemons. For this very purpose there are two config knobs: nfacctd_templates_receiver (to export received templates) and nfacctd_templates_port (to define a port to receive templates from the rest of the cluster). When multiple daemons are clustered a replicator (nfacctd with 'tee' plugin configured is required). Here is an example of the configs needed: nfacctd clustered daemon config (other daemons in the same cluster would be listening on different nfacctd_templates_port ports, ie. 20002, 20003, etc.): ! ! [ .. existing config .. ] ! nfacctd_templates_receiver: 127.0.0.1:10000 nfacctd_templates_port: 20001 The replicator piece needs two bits, a config and a tee_receivers file (in the example receivers.lst). Here is the config: nfacctd_port: 10000 ! plugins: tee[a] ! tee_receivers[a]: /path/to/receivers.lst tee_transparent[a]: true And here is the receivers.lst file: id=1 ip=127.0.0.1:20001,127.0.0.1:20002,127.0.0.1:20003,[..] Vc. Examples configuring NetFlow v9/IPFIX export Example to configure NetFlow v9 export on a Cisco running IOS/IOS-XE: ip flow-cache timeout active 1 ip flow-cache mpls label-positions 1 ! ip flow-export source Loopback0 ip flow-export version 9 bgp-nexthop ip flow-export template timeout-rate 1 ip flow-export template refresh-rate 4 ip flow-export destination X.X.X.X 2100 ! interface GigabitEthernet0/0 ip address Y.Y.Y.Y Z.Z.Z.Z ip flow ingress Example to configure NetFlow v9 export on a Cisco running IOS-XR: sampler-map NFACCTD-SMP random 1 out-of 10 ! flow monitor-map NFACCTD-MON record ipv4 exporter NFACCTD-EXP ! flow exporter-map NFACCTD-EXP version v9 transport udp 2100 destination X.X.X.X ! interface GigabitEthernet0/0/0/1 ipv4 address Y.Y.Y.Y Z.Z.Z.Z flow ipv4 monitor NFACCTD-MON sampler NFACCTD-SMP ingress Example to configure IPFIX export on a Juniper: services { flow-monitoring { version-ipfix { template ipv4 { flow-active-timeout 60; flow-inactive-timeout 70; template-refresh-rate seconds 30; option-refresh-rate seconds 30; ipv4-template; } } } } chassis { fpc 0 { sampling-instance s1; } } forwarding-options { sampling { instance { s1 { input { rate 10; } family inet { output { flow-server X.X.X.X { port 2100; version-ipfix { template { ipv4; } } } inline-jflow { source-address Y.Y.Y.Y; } } } } } } } Example to configure NetFlow v9 export on a Huawei: ip netstream timeout active 1 ip netstream timeout inactive 5 ip netstream mpls-aware label-and-ip ip netstream export version 9 origin-as bgp-nexthop ip netstream export index-switch 32 ip netstream export template timeout-rate 1 ip netstream sampler fix-packets 1000 inbound ip netstream export source Y.Y.Y.Y ip netstream export host X.X.X.X 2100 ipv6 netstream timeout active 1 ipv6 netstream timeout inactive 5 ipv6 netstream mpls-aware label-and-ip ipv6 netstream export version 9 origin-as bgp-nexthop ipv6 netstream export index-switch 32 ipv6 netstream export template timeout-rate 1 ipv6 netstream sampler fix-packets 1000 inbound interface Eth-Trunk1.100 ip netstream inbound ipv6 netstream inbound Contribution of further configuration examples for Cisco and Juniper devices and/or other relevant vendors is more than welcome. VI. Running the NFLOG-based daemon (uacctd) All examples about pmacctd are also valid for uacctd with the exception of directives that apply exclusively to libpcap. If you've skipped examples in the "Running the libpcap-based daemon (pmacctd)" section, please read them before continuing. All configuration keys available are in the CONFIG-KEYS document. The daemon depends on the package libnetfilter-log-dev in Debian/Ubuntu, libnetfilter_log in CentOS/RHEL (or equivalent package in the preferred Linux distribution). The support for NFLOG is disabled by default and should be enabled as follows: shell> ./configure --enable-nflog NFLOG_CFLAGS and NFLOG_LIBS can be used if includes and library are not in default locations. The Linux NFLOG infrastructure requires a couple parameters in order to work properly: the NFLOG multicast group (uacctd_group) to which captured packets have to be sent to and the Netlink buffer size (uacctd_nl_size). The default buffer settings (128KB) typically works OK for small environments. The traffic is captured with an iptables rule. For example in one of the following ways: * iptables -t mangle -I POSTROUTING -j NFLOG --nflog-group 5 * iptables -t raw -I PREROUTING -j NFLOG --nflog-group 5 Apart from determining how and what traffic to capture with iptables, which is topic outside the scope of this document, the most relevant point is the "--nflog-nlgroup" iptables setting has to match with the "uacctd_group" uacctd one. To review the packet flow in iptables: https://commons.wikimedia.org/wiki/File:Netfilter-packet-flow.svg A couple examples follow: Run uacctd reading configuration from a specified file. shell> uacctd -f uacctd.conf Daemonize the process; aggregate data by sum_host (by host, summing inbound + outbound traffic); write to a local MySQL server. Listen on NFLOG multicast group #5. Let's make pmacct divide data into historical time-bins of 5 minutes. Let's disable UPDATE queries and hence align refresh time with the timeslot length. Finally, let's make use of a SQL table, version 4: ! uacctd_group: 5 daemonize: true plugins: mysql aggregate: sum_host sql_refresh_time: 300 sql_history: 5m sql_history_roundoff: mh sql_table_version: 4 sql_dont_try_update: true ! ... VII. Running the pmacct IMT client (pmacct) The 'pmacct' client tool allows to query memory tables. Messaging happens over a UNIX pipe file: authorization is strictly connected to permissions of the pipe file. Note: while writing queries commandline, it may happen to write chars with a special meaning for the shell itself (ie. ; or *). Mind to either escape ( \; or \* ) them or put in quotes ( " ). Show all available pmacct client commandline switches: shell> pmacct -h Fetch data stored in the memory table: shell> pmacct -s Fetch data stored in the memory table using JSON output (and nicely format it with the 'jq' tool): shell> pmacct -s -O json | jq Match data between source IP 192.168.0.10 and destination IP 192.168.0.3 and return a formatted output; display all fields (-a), this way the output is easy to be parsed by tools like awk/sed; each unused field will be zero-filled: shell> pmacct -c src_host,dst_host -M 192.168.0.10,192.168.0.3 -a Similar to the previous example; it is requested to reset data for matched entries; the server will return the actual counters to the client, then will reset them: shell> pmacct -c src_host,dst_host -M 192.168.0.10,192.168.0.3 -r Fetch data for IP address dst_host 10.0.1.200; we also ask for a 'counter only' output ('-N') suitable, this time, for injecting data in tools like MRTG or RRDtool (sample scripts are in the examples/ tree). Bytes counter will be returned (but the '-n' switch allows also select which counter to display). If multiple entries match the request (ie because the query is based on dst_host but the daemon is actually aggregating traffic as "src_host, dst_host") their counters will be summed: shell> pmacct -c dst_host -N 10.0.1.200 Query the memory table available via pipe file /tmp/pipe.eth0: shell> pmacct -c sum_port -N 80 -p /tmp/pipe.eth0 Find all data matching host 192.168.84.133 as either their source or destination address. In particular, this example shows how to use wildcards and how to spawn multiple queries (each separated by the ';' symbol). Take care to follow the same order when specifying the primitive name (-c) and its actual value ('-M' or '-N'): shell> pmacct -c src_host,dst_host -N "192.168.84.133,*;*,192.168.84.133" Find all web and smtp traffic; we are interested in have just the total of such traffic (for example, to split legal network usage from the total); the output will be a unique counter, sum of the partial (coming from each query) values. shell> pmacct -c src_port,dst_port -N "25,*;*,25;80,*;*,80" -S Show traffic between the specified hosts; this aims to be a simple example of a batch query; note that as value of both '-N' and '-M' switches it can be supplied a value like: 'file:/home/paolo/queries.list': actual values will be read from the specified file (and they need to be written into it, one per line) instead of commandline: shell> pmacct -c src_host,dst_host -N "10.0.0.10,10.0.0.1;10.0.0.9,10.0.0.1;10.0.0.8,10.0.0.1" shell> pmacct -c src_host,dst_host -N "file:/home/paolo/queries.list" VIII. Running the RabbitMQ/AMQP plugin The Advanced Message Queuing Protocol (AMQP) is an open standard for passing business messages between applications. RabbitMQ is a messaging broker, an intermediary for messaging, which implements AMQP. pmacct RabbitMQ/AMQP plugin is designed to send aggregated network traffic data, in JSON or Avro format, through a RabbitMQ server to 3rd party applications (typically, but not limited to, noSQL databases like ElasticSearch, InfluxDB, etc.). Requirements to use the plugin are: * A working RabbitMQ server: http://www.rabbitmq.com/ * RabbitMQ C API, rabbitmq-c: https://github.com/alanxz/rabbitmq-c/ * Libjansson to cook JSON objects: http://www.digip.org/jansson/ Additionally, the Apache Avro C library (http://avro.apache.org/) needs to be installed to be able to send messages packed using Avro (you will also need to pass --enable-avro to the configuration script). Once these elements are installed, pmacct can be configured for compiling. pmacct makes use of pkg-config for finding libraries and headers location and checks some "typical" default locations, ie. /usr/local/lib and /usr/local/include. So all you should do is just: shell> ./configure --enable-rabbitmq --enable-jansson But, for example, should you have installed RabbitMQ in /usr/local/rabbitmq and pkg-config is unable to help, you can supply this non-default location as follows (assuming you are running the bash shell): shell> export RABBITMQ_LIBS="-L/usr/local/rabbitmq/lib -lrabbitmq" shell> export RABBITMQ_CFLAGS="-I/usr/local/rabbitmq/include" shell> ./configure --enable-rabbitmq --enable-jansson You can check further information on how to compile pmacct with JSON/libjansson support in the section "Compiling pmacct with JSON support" of this document. You can check further information on how to compile pmacct with Avro support in the section "Compiling pmacct with Apache Avro support" of this document. Then "make; make install" as usual. Following a configuration snippet showing a basic RabbitMQ/AMQP plugin configuration (assumes: RabbitMQ server is available at localhost; look all configurable directives up in the CONFIG-KEYS document): ! .. plugins: amqp ! aggregate: src_host, dst_host, src_port, dst_port, proto, tos amqp_output: json amqp_exchange: pmacct amqp_routing_key: acct amqp_refresh_time: 300 amqp_history: 5m amqp_history_roundoff: m ! .. pmacct will only declare a message exchange and provide a routing key, ie. it will not get involved with queues at all. A basic consumer script, in Python, is provided as sample to: declare a queue, bind the queue to the exchange and show consumed data on the screen or post to a REST API. The script is located in the pmacct default distribution tarball in 'examples/amqp/amqp_receiver.py' and requires the 'pika' Python module installed. Should this not be available, installation instructions are available at the following page: http://www.rabbitmq.com/tutorials/tutorial-one-python.html IX. Running the Kafka plugin Apache Kafka is a distributed streaming platform. Its qualities being: fast, scalable, durable and distributed by design. pmacct Kafka plugin is designed to send aggregated network traffic data, in JSON or Avro format, through a Kafka broker to 3rd party applications (typically, but not limited to, noSQL databases like ElasticSearch, InfluxDB, etc.). Requirements to use the plugin are: * A working Kafka broker (and Zookeeper server): http://kafka.apache.org/ * Librdkafka: https://github.com/edenhill/librdkafka/ * Libjansson to cook JSON objects: http://www.digip.org/jansson/ Additionally, the Apache Avro C library (http://avro.apache.org/) needs to be installed to be able to send messages packed using Avro (you will also need to pass --enable-avro to the configuration script). Once these elements are installed, pmacct can be configured for compiling. pmacct makes use of pkg-config for finding libraries and headers location and checks some default locations, ie. /usr/local/lib and /usr/local/include. If this is satisfactory, all you should do is just: shell> ./configure --enable-kafka --enable-jansson But, for example, should you have installed Kafka in /usr/local/kafka and pkg- config is unable to help, you can supply this non-default location as follows (assuming you are running the bash shell): shell> export KAFKA_LIBS="-L/usr/local/kafka/lib -lrdkafka" shell> export KAFKA_CFLAGS="-I/usr/local/kafka/include" shell> ./configure --enable-kafka --enable-jansson You can check further information on how to compile pmacct with JSON/libjansson support in the section "Compiling pmacct with JSON support" of this document. As a proof-of-concept once some data is produced in JSON format to a Kafka topic, it can be consumed with the kafka-console-consumer tool, part of standard Kafka distribution, as: kafka-console-consumer --bootstrap-server \ --topic You can check further information on how to compile pmacct with Avro support in the section "Compiling pmacct with Apache Avro support" of this document. Also, if using Confluent Platform, a Schema Registry component is included with it and allows for seamless inter-change of Avro schemas among producers and consumers of a Kafka broker; to make use of this, pmacct should be compiled with libserdes support: shell> ./configure --enable-kafka --enable-jansson --enable-avro --enable-serdes For further info on Confluent Schema Registry and libserdes: https://docs.confluent.io/current/schema-registry/docs/index.html https://github.com/confluentinc/libserdes NOTE: issues have been reported for certain cases when using libserdes on a system with OpenSSL 1.0.x installed due to some interactions with libcurl. Please upgrade to OpenSSL 1.1.x or greater. Then "make; make install" as usual. As a proof-of-concept, when the Schema Registry is in use, once some data is produced in Avro format to a Kafka topic, it can be consumed with the kafka-avro-console-consumer tool, part of standard Confluent Platform distribution, as: kafka-avro-console-consumer --bootstrap-server \ --topic \ --property schema.registry.url= \ --skip-message-on-error Following a configuration snippet showing a basic Kafka plugin configuration (assumes: Kafka broker is available at 127.0.0.1 on port 9092; look all kafka_* configurable directives up in the CONFIG-KEYS document): ! .. plugins: kafka ! aggregate: src_host, dst_host, src_port, dst_port, proto, tos kafka_output: json kafka_topic: pmacct.acct ! kafka_broker_host: 127.0.0.1 kafka_refresh_time: 300 kafka_history: 5m kafka_history_roundoff: m ! ! [Optional] librdkafka config for buffer tuning, ssl config, etc. ! kafka_config_file: /path/to/librdkafka.conf ! ! [Optional] Schema Registry URL ! kafka_avro_schema_registry: https://localhost ! .. A basic script, in Python, is provided as sample to consume JSON/Avro data from a given topic and perform some simple actions (ie. post data to a new Kafka topic, post to a URL or print to screen). The script is located in the pmacct distribution code in 'examples/kafka/kafka_consumer.py' and requires the confluent-kafka-python Python module installed (among the others, check script header). Should this not be available you can read on the following page how to get it installed: https://github.com/confluentinc/confluent-kafka-python This is a pointer to the quick start guide to Kafka: https://kafka.apache.org/quickstart IXa. Notes on writing to Kafka 1) When using Kafka over a dedicated node or VM, you may have to update the default Kafka server configuration. Edit the file named server.properties under the config folder of your kafka installation. Uncomment the following parameters: * listeners, * advertised.listeners, * listener.security.protocol.map and configure it according to your Kafka design. Taking a simple example where there is one single Kafka node used for both Zookeeper and Kafka and this node is using and ip address like 172.16.2.1. Those three parameters will look like this: listeners=PLAINTEXT://172.16.2.1:9092 advertised.listeners=PLAINTEXT://172.16.2.1:9092 listener.security.protocol.map=PLAINTEXT:PLAINTEXT,SSL:SSL,SASL_PLAINTEXT:SASL_PLAINTEXT,SASL_SSL:SASL_SSL 2) When the amount of data published to Kafka is substantial, ie. in the order of thousands of entries per second, batching (MessageSet) is needed in order to avoid every single object being produced singularly to Kafka brokers. Two strategies are available for this a) pmacct batching via the kafka_multi_values feature (legacy) or b) librdkafka batching (recommended) via, as per the lib documentation, "The two most important configuration properties for performance tuning [are]: * batch.num.messages : the minimum number of messages to wait for to accumulate in the local queue before sending off a message set. * queue.buffering.max.ms : how long to wait for batch.num.messages to fill up in the local queue.". Note: up until librdkafka 0.11, the default was kept very low to favour low-latency. A sensible minimum value in high-throughput scenarios is 50ms. Also, intuitively, queue.buffering.max.messages, the "Maximum number of messages allowed on the producer queue", should be kept greater than the batch.num.messages. These knobs can all be supplied to librdkafka through pmacct via a file pointed by kafka_config_file, as global settings, ie.: global, queue.buffering.max.messages, 8000000 global, batch.num.messages, 100000 global, queue.buffering.max.ms, 50 3) Configuring statistics.interval.ms to a positive value in kafka_config_file will make librdkafka log plenty of internal metrics - as documented at the following wiki page: https://github.com/edenhill/librdkafka/wiki/Statistics For example: global, statistics.interval.ms, 60000 This can come especially useful to perform more specific health checks (ie. is the daemon as a whole successfully processing data?) against the running daemon beyond checking, for example, pid files. 4) SSL and SASL supports in librdkafka are totally configuration based, meaning they can be enabled via the kafka_config_file. More info is available at the pages: https://github.com/edenhill/librdkafka/wiki/Using-SSL-with-librdkafka and https://github.com/edenhill/librdkafka/wiki/Using-SASL-with-librdkafka . For example, quoting the librdkafka page for SSL: global, security.protocol, ssl global, ssl.ca.location, ca-cert global, ssl.certificate.location, client_?????_client.pem global, ssl.key.location, client_?????_client.key global, ssl.key.password, abcdefgh Example: connect to a Confluent Cloud instance via SASL + SSL: global, bootstrap.servers, .aws.confluent.cloud:9092 global, security.protocol, SASL_SSL global, sasl.mechanisms, PLAIN global, sasl.username, global, sasl.password, global, session.timeout.ms, 45000 Example: connect to Microsoft Azure via the Kafka API using SASL: global, security.protocol, SASL_SSL global, sasl.mechanism, PLAIN global, sasl.username, $ConnectionString global, sasl.password, Endpoint=sb://XXX.windows.net/;SharedAccessKeyName=YYY;SharedAccessKey=ZZZ;EntityPath=JJJ 5) When integrating pmacct with the PNDA framework ( http://www.pnda.io/ ), a guide on how to make the two communicate has been written-up and published at this URL: https://github.com/jbotello7381/pmacct_to_pnda.io . X. Internal buffering and queueing Two options are provided for internal buffering and queueing: 1) a home-grown circular queue implementation available since day one of pmacct (configured via plugin_pipe_size and documented in docs/INTERNALS) and 2) a ZeroMQ queue (configured via plugin_pipe_zmq and plugin_pipe_zmq_* directives). The ZeroMQ queue is the recommended one. For a quick comparison: while relying on a ZeroMQ queue does introduce an external dependency, ie. libzmq, it reduces the amount of trial and error needed to fine tune plugin_buffer_size and plugin_pipe_size directives needed by the home-grown queue implementation. The home-grown circular queue has no external dependencies and is configured, for example, as: plugins: print[blabla] plugin_buffer_size[blabla]: 10240 plugin_pipe_size[blabla]: 1024000 For more information about the home-grown circular queue, consult plugin_buffer_size and plugin_pipe_size entries in CONFIG-KEYS and docs/INTERNALS "Communications between core process and plugins" chapter. ZeroMQ, from 0MQ The Guide, "looks like an embeddable networking library but acts like a concurrency framework. It gives you sockets that carry atomic messages across various transports like in-process, inter-process, TCP, and multicast. You can connect sockets N-to-N with patterns like fan-out, pub-sub, task distribution, and request-reply. It's fast enough to be the fabric for clustered products. Its asynchronous I/O model gives you scalable multicore applications, built as asynchronous message-processing tasks. [ .. ]". pmacct integrates ZeroMQ using a pub-sub queue architecture, using ephemeral TCP ports and implementing plain authentication (username and password, auto-generated at runtime). The only requirement to use a ZeroMQ queue is to have the latest available stable release of libzmq installed on the system (http://zeromq.org/intro:get-the-software , https://github.com/zeromq/libzmq/releases). Once this is installed, pmacct can be configured for compiling. pmacct makes use of pkg-config for finding libraries and headers location and checks some "typical" default locations, ie. /usr/local/lib and /usr/local/include. So all you should do is just: shell> ./configure --enable-zmq But, for example, should you have installed ZeroMQ in /usr/local/zeromq and should also pkg-config be unable to help, the non-default location can be supplied as follows (bash shell assumed): shell> export ZMQ_LIBS="-L/usr/local/zeromq/lib -lzmq" shell> export ZMQ_CFLAGS="-I/usr/local/zeromq/include" shell> ./configure --enable-zmq Then "make; make install" as usual. Following a configuration snippet showing how easy is to leverage ZeroMQ for queueing (see CONFIG-KEYS for all ZeroMQ-related options): plugins: print[blabla] plugin_pipe_zmq[blabla]: true plugin_pipe_zmq_profile[blabla]: micro Please review the standard buffer profiles, plugin_pipe_zmq_profile, in CONFIG-KEYS; Q21 of FAQS describes how to estimate the amount of flows/samples per second of your deployment. XI. Quickstart guide to packet/flow classification Packet classification is a feature available for pmacctd (libpcap-based daemon), uacctd (NFLOG-based daemon), nfacctd (NetFlow daemon) via NetFlow v9/IPFIX IE #315 (dataLinkFrameSection) and sfacctd (sFlow daemon) via sFlow v5 raw header sample and the current approach is to leverage the popular free, open-source nDPI library. To enable the feature please follow these steps: 1) Download pmacct from its webpage (http://www.pmacct.net/) or from its GitHub repository (https://github.com/pmacct/pmacct). 2) Download nDPI from its GitHub repository (https://github.com/ntop/nDPI). The nDPI API tends to change frequently in a non-backward compatible fashion. This means that not every version of pmacct will work with any version of nDPI. pmacct 1.7.8 works against nDPI 4.0-stable, 4.2-stable and 4.4-stable. pmacct 1.7.9 does work against 4.6-stable. pmacct 1.8.0 dows work against 4.14-stable. 3) Configure for compiling, compile and install the downloaded nDPI library, ie. inside the nDPI directory: shell> ./autogen.sh; ./configure; make; make install; ldconfig 4) Configure for compiling pmacct with the --enable-ndpi switch. Then compile and install, ie.: If downloading a release from http://www.pmacct.net , from inside the pmacct directory: shell> ./configure --enable-ndpi; make; make install If downloading code from https://github.com/pmacct/pmacct , from inside the pmacct directory: shell> ./autogen.sh; ./configure --enable-ndpi; make; make install If using a nDPI library that is not installed (or not installed in a default location) on the system, then NDPI_LIBS and NDPI_CFLAGS should be set to the location where nDPI headers and dynamic library are lying, for example: shell> NDPI_LIBS="-L/path/to/nDPI/src/lib/.libs -lndpi" shell> NDPI_CFLAGS=-I/path/to/nDPI/src/include shell> export NDPI_LIBS NDPI_CFLAGS shell> ./configure --enable-ndpi shell> make; make install 5) Configure pmacct. The following sample configuration is based on pmacctd and the print plugin with formatted output to stdout: daemonize: true pcap_interface: eth0 snaplen: 700 ! plugins: print ! aggregate: src_host, dst_host, src_port, dst_port, proto, tos, class What enables packet classification is the use of the 'class' primitive as part of the supplied aggregation method. Further classification-related options, such as timers, attempts, etc., are documented in the CONFIG-KEYS document (classifier_* directives). 6) Execute pmacct as: shell> pmacctd -f /path/to/pmacctd.conf Flow classification is a feature available for nfacctd (NetFlow daemon) and relies on NetFlow v9/IPFIX IEs #94, #95 and #96. Where IE #95 (applicationId) is part of flow data and can be looked up against Options data where IE #96 (applicationName) and IE #94 (applicationDescription) are supplied - specifically applicationName is the one used for the 'class' primitive value. To enable the feature, if the NetFlow/ IPFIX exporter does follow Cisco NBAR2 / RFC6759 implementation (and pmacct nfprobe plugin does!), simply add the 'class' primitive to the existing 'aggregate' aggregation method. XII. Quickstart guide to setup a NetFlow/IPFIX agent/probe pmacct is able to export traffic data through both NetFlow and sFlow protocols. This section covers NetFlow/IPFIX and next one covers sFlow. While NetFlow v5 is fixed by nature, v9 adds (to v5) flexibility allowing to transport custom information (ie. classification information or custom tags to remote collectors) and IPFIX adds (to v9) the chance to transport private information (PEN) and variable-length elements. Below the configuration guide: a) usual initial steps: download pmacct, unpack it, compile it. b) build NetFlow probe configuration, using pmacctd: ! daemonize: true pcap_interface: eth0 aggregate: src_host, dst_host, src_port, dst_port, proto, tos plugins: nfprobe nfprobe_receiver: 192.168.1.1:2100 ! nfprobe_receiver: [FD00::2]:2100 nfprobe_version: 10 ! nfprobe_engine: 1:1 ! nfprobe_timeouts: tcp=120:maxlife=3600 ! ! networks_file: /path/to/networks.lst !... This is a basic working configuration. Additional probe features include: 1) generate ASNs by using a networks_file pointing to a valid Networks File (see examples/ directory) and adding src_as, dst_as primitives to the 'aggregate' directive; alternatively, it is possible to generate ASNs from the pmacctd BGP thread. The following fragment can be added to the config above: pmacctd_as: bgp bgp_daemon: true bgp_daemon_ip: 127.0.0.1 bgp_agent_map: /path/to/bgp_agent.map bgp_daemon_port: 17917 The bgp_daemon_port can be changed from the standard BGP port (179/TCP) in order to co-exist with other BGP routing software which might be running on the same host. Depending on the BGP implementation used, it is possible to peer each other this way using 127.0.0.1 as bgp_daemon_ip. It may however be necessary to use a different listening IP address for pmacctd BGP and the other BGP software (see below). In pmacctd, bgp_agent_map does the trick of mapping 0.0.0.0 to the IP address of the BGP peer (ie. 127.0.0.1: 'bgp_ip=127.0.0.1 ip=0.0.0.0'); this setup, while generic, was tested working in conjunction with Quagga, FRR and BIRD. Example fragment of the FRR configuration: router bgp Y bgp router-id X.X.X.X bgp ebgp-requires-policy bgp log-neighbor-changes neighbor 127.0.0.1 remote-as Y neighbor 127.0.0.1 port 17917 neighbor 127.0.0.1 update-source 127.0.0.2 ! address-family ipv4 unicast neighbor 127.0.0.1 activate neighbor 127.0.0.1 route-map IMPORT-pmacct-4 in neighbor 127.0.0.1 route-map EXPORT-pmacct-4 out ! address-family ipv6 unicast neighbor 127.0.0.1 activate neighbor 127.0.0.1 route-map IMPORT-pmacct-6 in neighbor 127.0.0.1 route-map EXPORT-pmacct-6 out ! route-map IMPORT-pmacct-4 deny 10 route-map EXPORT-pmacct-4 permit 10 route-map IMPORT-pmacct-6 deny 10 route-map EXPORT-pmacct-6 permit 10 ! NOTE: FRR supports peering with a different implementation on the same system as of February 2022 (> 8.1.0), with the following caveats: * FRR's 'bgpd_options' directive must be configured to listen on a different IP address than pmacctd's bgp_daemon_ip. * FRR cannot initiate BGP sessions with update-source 127.0.0.1 or ::1. Setting up a 127.0.0.2 address on the system and binding FRR's bgpd to it allows the BGP session to come up fine. Example fragment of the BIRD configuration: protocol bgp pmacctd46 { description "pmacctd"; local 127.0.0.1 as YYYYY; neighbor 127.0.0.2 port 179 as YYYYY; rr client; hold time 90; keepalive time 30; graceful restart; ipv4 { next hop self; import filter { reject; }; export filter { accept; }; }; ipv6 { next hop address 127.0.0.1; import filter { reject; }; export filter { accept; }; }; } NOTE: Bird Internet Routing Daemon seems to dislike peerings on the same IP address it is configured on, even though the remote peer listens on a different TCP port; if Bird listens on 127.0.0.1 then setting up a 127.0.0.2 address on the system and binding pmacctd to it allows the BGP session to come up fine in a 127.0.0.1 <-> 127.0.0.2 fashion. 2) encode flow classification information in NetFlow v9 / IPFIX like Cisco does with its NBAR2 implementation / RFC6759. This can be done by introducing the 'class' primitive to the afore mentioned 'aggregate' and add the extra config directive: aggregate: class, src_host, dst_host, src_port, dst_port, proto, tos snaplen: 700 Further information on this topic can be found in the 'Quickstart guide to packet classification' section of this document. 3) add direction (ingress, egress) awareness to measured IP traffic flows. Direction can be defined statically (in, out) or inferred dynamically (tag, tag2) via the use of the nfprobe_direction directive. Let's look at a dynamic example using tag2; first, add the following lines to the daemon configuration: nfprobe_direction[plugin_name]: tag2 pre_tag_map: /path/to/pretag.map then edit the tag map as follows. A return value of '1' means ingress while '2' is translated to egress. It is possible to define L2 and/or L3 addresses to recognize flow directions. The 'set_tag2' primitive (tag2) will be used to carry the return value: set_tag2=1 filter='dst host XXX.XXX.XXX.XXX' set_tag2=2 filter='src host XXX.XXX.XXX.XXX' set_tag2=1 filter='ether src XX:XX:XX:XX:XX:XX' set_tag2=2 filter='ether dst XX:XX:XX:XX:XX:XX' Indeed in such a case, the 'set_tag' primitive (tag) can be leveraged to other uses (ie. filter sub-set of the traffic for flow export); 4) add interface (input, output) awareness to measured IP traffic flows. Interfaces can be defined only in addition to direction. Interface can be either defined statically (<1-4294967295>) or inferred dynamically (tag, tag2) with the use of the nfprobe_ifindex directive. Let's look at a dynamic example using tag; first add the following lines to the daemon config: nfprobe_direction[plugin_name]: tag nfprobe_ifindex[plugin_name]: tag2 pre_tag_map: /path/to/pretag.map then edit the tag map as follows: set_tag=1 filter='dst net XXX.XXX.XXX.XXX/WW' jeq=eval_ifindexes set_tag=2 filter='src net XXX.XXX.XXX.XXX/WW' jeq=eval_ifindexes set_tag=1 filter='dst net YYY.YYY.YYY.YYY/ZZ' jeq=eval_ifindexes set_tag=2 filter='src net YYY.YYY.YYY.YYY/ZZ' jeq=eval_ifindexes set_tag=1 filter='ether src YY:YY:YY:YY:YY:YY' jeq=eval_ifindexes set_tag=2 filter='ether dst YY:YY:YY:YY:YY:YY' jeq=eval_ifindexes set_tag=999 filter='net 0.0.0.0/0' ! set_tag2=100 filter='dst host XXX.XXX.XXX.XXX' label=eval_ifindexes set_tag2=100 filter='src host XXX.XXX.XXX.XXX' set_tag2=200 filter='dst host YYY.YYY.YYY.YYY' set_tag2=200 filter='src host YYY.YYY.YYY.YYY' set_tag2=200 filter='ether src YY:YY:YY:YY:YY:YY' set_tag2=200 filter='ether dst YY:YY:YY:YY:YY:YY' The set_tag=999 works as a catch all for undefined L2/L3 addresses so to prevent searching further in the map. In the example above direction is set first then, if found, interfaces are set, using the jeq/label pre_tag_map construct. c) build NetFlow collector configuration, using nfacctd: ! daemonize: true nfacctd_ip: 192.168.1.1 nfacctd_port: 2100 plugins: memory[display] aggregate[display]: src_host, dst_host, src_port, dst_port, proto ! aggregate[display]: class, src_host, dst_host, src_port, dst_port, proto d) Ok, we are done ! Now fire both daemons: shell a> pmacctd -f /path/to/configuration/pmacctd-nfprobe.conf shell b> nfacctd -f /path/to/configuration/nfacctd-memory.conf XIII. Quickstart guide to setup a sFlow agent/probe pmacct can export traffic data via sFlow; such protocol is different from NetFlow/ IPFIX: in short, it works by sampling individual packets (and exporting portions of them) instead of caching uni-directional flows as it happens for NetFlow; such a stateless approach makes sFlow a light export protocol well-tailored for high-speed networks. Furthermore, sFlow v5 can be extended much like NetFlow v9/IPFIX: meaning classification information (if nDPI is compiled in, see 'Quickstart guide to packet classification' section of this document), tags and/or basic Extended Gateway info (ie. src_as, dst_as) can be easily included in the record structure being exported. For classification info to be included in sFlow packets please make sure to specify the classifier_num_roots directive in the config. Here a quickstarter sFlow probe configuration, using pmacctd: ! daemonize: true pcap_interface: eth0 plugins: sfprobe sampling_rate: 20 sfprobe_agentsubid: 1402 sfprobe_receiver: 192.168.1.1:6343 ! ! sfprobe_agentip: FD00::1 ! sfprobe_source_ip: FD00::1 ! sfprobe_receiver: [FD00::2]:6343 ! ! networks_file: /path/to/networks.lst ! snaplen: 700 ! ! classifier_num_roots: 512 ! !... XIV. Quickstart guide to setup the BGP daemon BGP can be run as a stand-alone collector daemon (pmbgpd) or as a thread within one of the traffic accounting daemons (ie. nfacctd). The stand-alone daemon is suitable for the Looking-Glass use-case, building BGP xconnects and consuming BGP data, real-time or at regular intervals; the thread solution is suitable for correlation of BGP with other data sources, ie. NetFlow, IPFIX, sFlow, etc. The thread implementation idea is to receive data-plane information, ie. via NetFlow, sFlow, etc., and control plane information, ie. full routing tables via BGP, from edge routers. Per-peer BGP RIBs are maintained to ensure local views of the network, a behaviour close to that of a BGP route-server. In case of routers with default-only or partial BGP views, the default route can be followed up (bgp_default_follow); also it might be desirable in certain situations, for example trading-off resources to accuracy, to map one or a set of flow agents to a BGP peer (bgp_agent_map). The following configuration snippet shows how to set up a BGP thread (ie. part of the NetFlow/IPFIX collector, nfacctd) which will bind to an IP address and will support up to a maximum number of 100 peers. Once PE routers start sending flow telemetry data and peer up, it should be possible to see the BGP-related fields, ie. as_path, peer_as_dst, local_pref, med, etc., correctly populated while querying the memory table: bgp_daemon: true bgp_daemon_ip: X.X.X.X bgp_daemon_max_peers: 100 ! bgp_daemon_as: 65555 nfacctd_as: bgp [ ... ] plugins: memory aggregate: src_as, dst_as, local_pref, med, as_path, peer_dst_as Setting up the stand-alone BGP collector daemon, pmbgpd, is not very different at all from the configuration above: bgp_daemon_ip: X.X.X.X bgp_daemon_max_peers: 100 ! bgp_daemon_as: 65555 bgp_table_dump_file: /path/to/spool/bgp-$peer_src_ip-%H%M.log bgp_table_dump_refresh_time: 300 Essentially: the 'bgp_daemon: true' line is not required and there is no need to instantiate plugins. On the other hand, the BGP daemon is instructed to dump BGP tables to disk every 300 secs with file names embedding the BGP peer info ($peer_src_ip) and time reference (%H%M). The BGP implementation, by default, reads the remote ASN upon receipt of a BGP OPEN message and dynamically presents itself as part of the same ASN - this is to ensure an iBGP relationship is established, even in multi ASN scenarios. It is possible to put pmacct in a specific ASN of choice by using the bgp_daemon_as configuration directive, for example, to establish an eBGP kind of relationship. Also, the daemon acts as a passive BGP neighbor and hence will never try to re-establish a fallen peering session. For debugging purposes related to the BGP feed(s), bgp_daemon_msglog_* configuration directives can be enabled in order to log BGP messaging. NOTE: especially when connecting large amounts of BGP peers, say hundreds or more, the almost synchronized attempt to connect from all of them may lead the kernel into a suspicion of TCP SYN Flood. While not recommending to disable TCP SYN flood protection, if slugginess is observed when taking up the BGP sessions and kernel logs indicate a possible SYN flooding, maybe it is good idea to tune and customize default kernel parameters. NOTE: currently if a BGP session is abruptly terminated by the remote peer, pmacct does rely on the underlying operating system to detect the condition (TCP keepalives) and close the socket. While this is going to change by detecting the condition inside the BGP daemon as a workaround TCP keepalive timers can be tuned to react in a reasonable amount of time. The example below is based on Linux and will disconnect a hung session in about two minutes: shell> sysctl -w net.ipv4.tcp_keepalive_time=60 shell> sysctl -w net.ipv4.tcp_keepalive_intvl=10 shell> sysctl -w net.ipv4.tcp_keepalive_probes=2 XIVa. Limiting AS-PATH and BGP community attributes length AS-PATH and BGP communities can get easily long when represented as strings. Sometimes only a small portion of their content is of interest and hence a simple filtering layer was developed in order to take special care of these BGP attributes. bgp_aspath_radius cuts the AS-PATH down after a specified amount of AS hops; whereas the bgp_stdcomm_pattern does a simple sub-string matching against standard BGP communities, filtering in only those that match (optionally, for better precision, a pre-defined number of characters can be wildcarded by employing the '.' symbol, like in regular expressions). See an example below: bgp_aspath_radius: 3 bgp_stdcomm_pattern: 12345: A detailed description of these configuration directives is, as usual, included in the CONFIG-KEYS document. XIVb. The source peer AS case The peer_src_as primitive adds useful insight in understanding where traffic enters the observed routing domain; but asymmetric routing impacts accuracy delivered by devices configured with either NetFlow or sFlow and the peer-as feature (as it only performs a reverse lookup, ie. a lookup on the source IP address, in the BGP table hence saying where it would route such traffic). pmacct offers a few ways to perform some mapping to tackle this issue and easily model both private and public peerings, both bi-lateral or multi-lateral. Find below how to use a map, reloadable at runtime, and its contents (for full syntax guide lines, please see the 'peers.map.example' file within the examples section): bgp_peer_src_as_type: map bgp_peer_src_as_map: /path/to/peers.map [/path/to/peers.map] set_tag=12345 ip=A.A.A.A in=10 bgp_nexthop=X.X.X.X set_tag=34567 ip=A.A.A.A in=10 set_tag=45678 ip=B.B.B.B in=20 src_mac=00:11:22:33:44:55 set_tag=56789 ip=B.B.B.B in=20 src_mac=00:22:33:44:55:66 Even though all this mapping is static, it can be auto-provisioned to a good degree by means of external scripts running at regular intervals and, for example, querying relevant routers via SNMP. In this sense, the bgpPeerTable MIB is a good starting point. Alternatively pmacct also offers the option to perform reverse BGP lookups. NOTES: * When using maps, the peer_src_as primitive should be used with care if egress NetFlow/IPFIX/sFlow is in use as it mainly relies on either the input interface index (ifIndex), the source MAC address, a reverse BGP next-hop lookup or a combination of these. For example, egress NetFlow enabled on edge devices on core interfaces may work as the input ifIndex would still be visible; egress NetFlow enabled on edge devices on edge interfaces may not work as, unless traffic is locally switched, it may be sampled upon exiting the observed netowrk. * "Source" MED, local preference, communities and AS-PATH have all been allocated aggregation primitives. Each carries its own peculiarities but the general concepts highlighed in this chapter apply to these aswell. Check CONFIG-KEYS out for the src_[med|local_pref|as_path|std_comm|ext_comm|lrg_comm]_[type|map] configuration directives. XIVc. Tracking entities on the own IP address space It might happen that not all entities connected to the observed network are running BGP but rather they get assigned provider IP prefixes redistributed into iBGP (different routing protocols, statics, directly connected, etc.). These can be private IP addresses or segments of the SP public address space. The common factor to all of them is that while being present in iBGP, these prefixes can't be distinguished any further due to the lack of attributes like AS-PATH or an ASN. One simple solution, although costly, could be to account on IP prefixes (other than BGP attributes). Another simple solution, especially if the enity is dominated by outbound traffic (ingress in the observed network) could be to use a bgp_peer_src_as_map directive, described previously (ie. make use of interface descriptions as a possible way to automate the process). Alternatively, bgp_stdcomm_pattern_to_asn and bgp_lrgcomm_pattern_to_asn config directives were developed to fit in this scenario: assuming procedures of a SP are (or can be changed) to label every relevant non-BGP speaking entity IP prefixes uniquely with BGP standard communities, this directive then allows to map the community to a peer AS/ origin AS couple as per the following example: XXXXX:YYYYY => Peer-AS=XXXXX, Origin-AS=YYYYY. XIVd. Preparing the router to BGP peer Once the collector is configured and started up the remaining step is to let routers export traffic samples to the collector and BGP peer with it. By configuring the same source IP address across both NetFlow and BGP features allows the collector to perform the required correlations without further configuration. Also, setting the BGP Router ID accordingly allows for more clear log messages. On the routers it is adviceable to configure the BGP session with the collector as a Route-Reflector (RR) client. A relevant configuration example for a Cisco IOS follows: router bgp 12345 neighbor X.X.X.X remote-as 12345 neighbor X.X.X.X update-source Loopback12345 neighbor X.X.X.X version 4 neighbor X.X.X.X send-community neighbor X.X.X.X route-reflector-client neighbor X.X.X.X description nfacctd A relevant configuration example for a Juniper router follows: protocols bgp { group rr-netflow { type internal; local-address Y.Y.Y.Y; family inet { any; } cluster Y.Y.Y.Y; neighbor X.X.X.X { description "nfacctd"; } } } A relevant configuration example for a IOS XR (Version 5.1.1) router follows: router bgp 12345 address-family ipv4 unicast additional-paths receive additional-paths send maximum-paths ibgp 6 additional-paths selection route-policy add_path ! af-group af-group-example address-family ipv4 unicast route-reflector-client ! session-group ibgp remote-as 12345 update-source Loopback0 ! neighbor-group ibgp-rr-client use session-group ibgp address-family ipv4 unicast use af-group af-group-example ! ! neighbor X.X.X.X use neighbor-group ibgp-rr-client ! ! Contribution of further configuration examples for Cisco and Juniper devices and/or other relevant vendors is more than welcome. In case ipv4 and ipv6 address-families are delivered over two distinct BGP sessions, typically one with ipv4 transport and one with ipv6 transport, then, since both ipv4 and ipv6 flows (NetFlow, IPFIX, sFlow) are typically delivered over a single ipv4 transport, a bgp_agent_map is needed in order to make the right correlation (see the 'filter' keyword in the example below). The map is defined in the daemon configuration ie. 'bgp_agent_map: /path/to/bgp_agent.map' then in the map itself for every router exporting flows and BGP peering the following lines should be defined: bgp_ip= ip= filter='ip or (vlan and ip)' bgp_ip= ip= filter='ip6 or (vlan and ip6)' XIVe. Example: writing flows augmented by BGP to a MySQL database The following setup is a realistic example for collecting an external traffic matrix to the ASN level (ie. no IP prefixes collected) for a MPLS-enabled IP carrier network. Samples are aggregated in a way which is suitable to get an overview of traffic trajectories, collecting much information where these enter the AS and where they get out. daemonize: true nfacctd_port: 2100 nfacctd_time_new: true plugins: mysql[5mins], mysql[hourly] sql_optimize_clauses: true sql_dont_try_update: true sql_multi_values: 1024000 sql_history_roundoff[5mins]: m sql_history[5mins]: 5m sql_refresh_time[5mins]: 300 sql_table[5mins]: acct_bgp_5mins sql_history_roundoff[hourly]: h sql_history[hourly]: 1h sql_refresh_time[hourly]: 3600 sql_table[hourly]: acct_bgp_1hr bgp_daemon: true bgp_daemon_ip: X.X.X.X bgp_daemon_max_peers: 100 bgp_aspath_radius: 3 bgp_follow_default: 1 nfacctd_as: bgp bgp_peer_src_as_type: map bgp_peer_src_as_map: /path/to/peers.map plugin_buffer_size: 10240 plugin_pipe_size: 1024000 aggregate: tag, src_as, dst_as, peer_src_as, peer_dst_as, peer_src_ip, peer_dst_ip, local_pref, as_path pre_tag_map: /path/to/pretag.map maps_refresh: true maps_entries: 3840 The content of the maps (bgp_peer_src_as_map, pre_tag_map) is meant to be pretty standard and will not be shown. As it can be grasped from the above configuration, the SQL schema was customized. Below a suggestion on how this can be modified for more efficiency - with additional INDEXes, to speed up specific queries response time, remaining to be worked out: create table acct_bgp_5mins ( id INT(4) UNSIGNED NOT NULL AUTO_INCREMENT, agent_id INT(4) UNSIGNED NOT NULL, as_src INT(4) UNSIGNED NOT NULL, as_dst INT(4) UNSIGNED NOT NULL, peer_as_src INT(4) UNSIGNED NOT NULL, peer_as_dst INT(4) UNSIGNED NOT NULL, peer_ip_src CHAR(15) NOT NULL, peer_ip_dst CHAR(15) NOT NULL, as_path CHAR(21) NOT NULL, local_pref INT(4) UNSIGNED NOT NULL, packets INT UNSIGNED NOT NULL, bytes BIGINT UNSIGNED NOT NULL, stamp_inserted DATETIME NOT NULL, stamp_updated DATETIME, PRIMARY KEY (id), INDEX ... ) TYPE=MyISAM AUTO_INCREMENT=1; create table acct_bgp_1hr ( id INT(4) UNSIGNED NOT NULL AUTO_INCREMENT, agent_id INT(4) UNSIGNED NOT NULL, as_src INT(4) UNSIGNED NOT NULL, as_dst INT(4) UNSIGNED NOT NULL, peer_as_src INT(4) UNSIGNED NOT NULL, peer_as_dst INT(4) UNSIGNED NOT NULL, peer_ip_src CHAR(15) NOT NULL, peer_ip_dst CHAR(15) NOT NULL, as_path CHAR(21) NOT NULL, local_pref INT(4) UNSIGNED NOT NULL, packets INT UNSIGNED NOT NULL, bytes BIGINT UNSIGNED NOT NULL, stamp_inserted DATETIME NOT NULL, stamp_updated DATETIME, PRIMARY KEY (id), INDEX ... ) TYPE=MyISAM AUTO_INCREMENT=1; Although table names are fixed in this example, ie. acct_bgp_5mins, it can be highly adviceable in real-life to run dynamic SQL tables, ie. table names that include time-related variables (see sql_table, sql_table_schema in CONFIG-KEYS). XIVf. Example: exporting BGP tables or messaging to files or AMQP/Kafka brokers. Both the stand-alone BGP collector daemon (pmbgpd) and the BGP thread within one of the traffic accounting daemons can: a) export/dump routing tables for all BGP peers at regular time intervals and b) log BGP messaging, real-time, with each of the BGP peers. Both features are useful for producing data useful for analytics, for troubleshooting and debugging. The former is beneficial to gain visibility in extra BGP data while providing event compression; the latter enables BGP analytics and BGP event management, for example spot unstable routes, trigger alarms on route hijacks, etc. Both features export data formatted as JSON or Avro messages. Please review the section "Compiling pmacct with JSON support" and, if planning to use Avro encoding, also the section "Compiling pmacct with Avro support". If writing to AMQP or Kafka brokers, compiling against RabbitMQ or Kafka libraries is required; read more in respectively the "Running the RabbitMQ/AMQP plugin" or "Running the Kafka plugin" sections of this document. If making use of the Schema Registry component of the Confluent Platform please also review the "Running the Kafka plugin" section. A basic dump of BGP tables at regular intervals (60 secs) to plain-text files, split by BGP peer and time of the day, is configured as follows: bgp_table_dump_file: /path/to/spool/bgp/bgp-$peer_src_ip-%H%M.txt bgp_table_dump_refresh_time: 60 A basic log of BGP messaging in near real-time to a plain-text file (which can be rotated by an external tool/script) is configured as follows: bgp_daemon_msglog_file: /path/to/spool/bgp/bgp-$peer_src_ip.log A basic dump of BGP tables at regular intervals (60 secs) to a Kafka broker, listening on the localhost and default port, is configured as follows: bgp_table_dump_kafka_topic: pmacct.bgp ! bgp_table_dump_kafka_broker_host: 127.0.0.1 bgp_table_dump_refresh_time: 60 ! ! [Optional] librdkafka config for buffer tuning, ssl config, etc. ! bgp_table_dump_kafka_config_file: /path/to/librdkafka.conf The equivalent bgp_table_dump_amqp_routing_key config directive can be used to make the above example work against a RabbitMQ broker. A basic log of BGP messaging in near real-time to a Kafka broker, listening on the localhost and default port, is configured as follows: bgp_daemon_msglog_kafka_topic: pmacct.bgp ! bgp_daemon_msglog_kafka_broker_host: 127.0.0.1 ! ! [Optional] librdkafka config for buffer tuning, ssl config, etc. ! bgp_daemon_msglog_kafka_config_file: /path/to/librdkafka.conf The equivalent bgp_daemon_msglog_amqp_routing_key config directive can be used to make the above example work against a RabbitMQ broker. A sample of both the BGP msglog and dump formats are captured in the following document: docs/MSGLOG_DUMP_FORMATS XIVg. Example: querying the BGP daemon via the Looking Glass server. The Looking Glass server allows to query RIBs stored in the BGP daemon. The server is asynchronous and uses ZeroMQ as transport layer to serve incoming queries. The LG server can be configured as follows: bgp_daemon_lg: true ! bgp_daemon_lg_user: pmacct ! bgp_daemon_lg_passwd: arealsmartpwd This configuration will enable plain authentication and will make the server listen on the localhost interface on port 17900/TCP. Following an example of querying with the Python client available in 'examples/lg': Get the list of available BGP peers: shell> examples/lg/pmbgp.py -g {"results": 1, "query_type": 2} {"peer_ip_src": "192.168.1.1", "peer_id": "192.168.1.1", "peer_tcp_port": 47248, "peer_as": 65500} Lookup IP prefix 192.168.2.0/24 for BGP peer 192.168.1.1: shell> examples/lg/pmbgp.py -a 192.168.2.0/24 -r 192.168.1.1 {"peer_ip_src": "192.168.1.1", "event_type": "lglass", "afi": 1, "safi": 1, "ip_prefix": "192.168.2.0/24", "bgp_nexthop": "192.168.1.254", "as_path": "65500 65501 65502", "comms": "65500:1 65500:100 65500:65500", "origin": 0, "local_pref": 100} Request/Reply Looking Glass formats are documented in 'docs/LOOKING_GLASS_FORMAT'. XIVh. Example: BGP Xconnects pmbgpd, the pmacct BGP standalone daemon, includes a BGP xconnect feature to map BGP peers (ie. routers) to BGP collectors (ie. nfacctd, sfacctd), acting essentially as a BGP proxy. While the proxying is non-transparent, BGP peers can be still identified by rheir BGP Router-ID. The aim of this feature is to facilitate operations, ie. point all BGP Peers to a single (or a couple) IP address(es) and facilitate re-sizing/re-balancing of the collection infrastructure without impacting (ie. re-configuring) BGP peers. A few caveats apply: 1) mapping works only against the IP source address and not the BGP Router ID, 2) only 1:1 relationships can be formed (ie. this is about cross-connecting sessions, not replication) and 3) only one session per BGP peer is supported (ie. multiple BGP agents are running on the same IP address or NAT traversal scenarios are not supported [yet]). The feature can be configured as follows: bgp_daemon_xconnect_map: /path/to/bgp_xconnects.map And the content of the map can be defined as follows: bgp_dst=10.0.0.1:17901 bgp_src=192.168.0.1 bgp_dst=10.0.0.1:17902 bgp_src=192.168.0.2 bgp_dst=10.0.0.1:17910 bgp_src=192.168.10.0/24 Where bgp_dst is the IPv4/IPv6 address and port of the BGP collector to cross- connect incoming (bgp_src) BGP session to and bgp_src is compared against the IPv4/IPv6 address (not BGP Router ID) of an incoming BGP session from a BGP peer. At the collector end one should set tmp_bgp_lookup_compare_ports (read more about this knob in CONFIG-KEYS) to true. In future BMP may be added as a transport for BGP data between pmbgpd and the collector (please get in touch if interested). An alternative to the use of pmbgpd is NAT/masquerading; in this context, pmbgpd brings the advantage of running fully in user-space. SO_REUSEPORT feature (read more about it in the section "Miscellaneous notes and troubleshooting tips" of this document) can be used in alternative to BGP xconnects, ie. to simplify the deployment on a single node, or in conjunction with BGP xconnects, ie. for massive multi-node scale-out. Leveraging SO_REUSEPORT, when used in alternative to BGP xconnects, may be best combined with a list of allowed IP addresses, ie. bgp_daemon_allow_file, to explicitly wire peers to collectors. When SO_REUSEPORT is used in conjunction with BGP xconnects there is virtually no config change to make (except perhaps the sensible decision to use different 'logfile' per pmbgpd daemon); below instead an example for SO_REUSEPORT in alternative to BGP xconnects where a few nfacctd daemons are deployed each running the BGP thread (indeed this is applicable to sfacctd as well): Daemon #1: allow-1.map: 10.0.0.0/25 nfacctd-1.conf: ... ! nfacctd_port: 2100 ! nfacctd_allow_file: /path/to/allow-1.map ! bgp_daemon: true bgp_daemon_port: 179 bgp_daemon_allow_file: /path/to/allow-1.map ! ... Daemon #2: allow-2.map: 10.0.0.128/25 nfacctd-2.conf: ... ! nfacctd_port: 2101 ! nfacctd_allow_file: /path/to/allow-2.map ! bgp_daemon: true bgp_daemon_port: 179 bgp_daemon_allow_file: /path/to/allow-2.map ! ... NOTE: when leveraging SO_REUSEPORT for TCP sessions in conjunction with allow files the kernel will try to naively balance the sessions over the listening daemons; this may take a few connection attempts before wiring the session to the right daemon; having an excessive numbers of daemons in the pool may slow down the convergence process. NOTE: each xconnect requires two file descriptors, one for the inbound TCP session, another for the outbound one. Popular default on Linux is 1024 descriptor per process, effectively limiting the amunt of sessions that can be xconnected to approx 500. Such default can be adjusted via 'ulimit -n'. For flow protocols, which are transported in UDP, a replication layer is needed in front, note the daemon listen on different nfacctd_port in the example above (read more on how to set up a replicator in the "Quickstart guide to setup a NetFlow/IPFIX/sFlow replicator" section of the document). nfacctd_allow_file is optional config key and, if created accordingly to bgp_daemon_allow_file, may ensure only flows from router for which a BGP sessions accepted should be processed. XIVi. Example: adding RPKI/ROA information Resource Public Key Infrastructure (RPKI) is a specialized public key infrastructure (PKI) framework designed to secure the Internet routing. RPKI uses certificates to allow Local Internet Registries (LIRs) to list the Internet number resources they hold. These attestations are called Route Origination Authorizations (ROAs). ROA information can be currently acquired in one of the two following ways: 1) it can be imported using the rpki_roas_file config directive from a file in the RIPE Validator or OpenBSD Validator formats, ie.: { "roas" : [ { "asn" : "AS2914", "prefix" : "128.121.0.0/16", "maxLength" : 16, "ta" : "ARIN" }, { "asn" : "AS2914", "prefix" : "128.121.0.0/19", "maxLength" : 24, "ta" : "ARIN" }, { "asn" : "AS2914", "prefix" : "128.121.4.32/29", "maxLength" : 29, "ta" : "ARIN" } ] } RPKI data in such format can be obtained for example at the following URL: https://rpki.gin.ntt.net/api/export.json . 2) Connecting to a RPKI RTR Cache for live ROA updates. Its IP address and port being defined by the rpki_rtr_cache config directive (a few more optional rpki_rtr_* directives are available and can be revived in the CONFIG-KEYS document). Once rpki_roas_file or rpki_rtr_cache are defined, traffic daemons (ie. nfacctd) and pmbgpd can be configured to correlate traffic, BGP and RPKI information. This is an example of how to configure nfacctd, the NetFlow/ IPFIX collector daemon: nfacctd_as: longest nfacctd_net: longest ! < .. > ! ! rpki_roas_file: /path/to/roas.json ! rpki_rtr_cache: 127.0.0.1:8323 ! rpki_rtr_cache_version: 1 ! bgp_daemon: true bgp_src_roa_type: bgp ! < .. > ! plugins: print[foo] ! aggregate[foo]: peer_src_ip, peer_dst_ip, src_host, dst_host, src_roa, dst_roa ! < .. > The ROA fields will be populated with one of these five values: 'u' Unknown, 'v' Valid, 'i' Invalid no overlaps, 'V' Invalid with a covering Valid prefix and 'U' Invalid with a covering Unknown prefix. XIVl. BGP daemon implementation concluding notes The implementation supports 4-bytes ASN, IPv4, IPv6, VPNv4 and VPNv6 (MP-BGP) address families and several IETF drafts and RFCs, ie. ADD-PATH, Shutdown Message and Large BGP Communities among the others; IPv4, IPv6 transports for BGP sessions are supported. When storing data via SQL, BGP primitives can be freely mix-and-matched with other primitives (ie. L2/L3/L4) when customizing the SQL table (sql_optimize_clauses: true). Environments making use of BGP Multi-Path should leverage ADD-PATH to advertise the multiple known paths; when correlating traffic flows to BGP, the correct BGP info is picked among the ones available using BGP next-hop (or for certain cases where traffic is sampled egress, IP next-hop if use_ip_next_hop is set to true) as selector; this does imply that BGP next-hop must be available in traffic data (ie. ADD-PATH selection is not compatible with NetFlow v5 which does not carry it) and BGP next-hop-self should not be set (otherwise all or multiple next-hops will just look like the same, hindering the decision process). TCP MD5 signature for BGP messages is also supported. For a review of all knobs and features see the CONFIG-KEYS document. XV. Quickstart guide to setup a NetFlow/IPFIX/sFlow replicator The 'tee' plugin is meant to replicate NetFlow/sFlow data to remote collectors. The plugin can act transparently, by preserving the original IP address of the datagrams, or as a proxy. The basic configuration of a replicator is very easy: all is needed is to define where to listen to for incoming packets, where to replicate them to and optionally a filtering layer, if required. Filtering bases on the standard pre_tag_map infrastructure; here will be presented only coarse-grained filtering against the NetFlow/sFlow source IP address (see next section for finer-grained filtering). nfacctd_port: 2100 nfacctd_ip: X.X.X.X ! plugins: tee[a], tee[b] tee_receivers[a]: /path/to/tee_receivers_a.lst tee_receivers[b]: /path/to/tee_receivers_b.lst ! tee_transparent: true ! ! pre_tag_map: /path/to/pretag.map ! plugin_buffer_size: 10240 plugin_pipe_size: 1024000 nfacctd_pipe_size: 1024000 An example of content of a tee_receivers map, ie. /path/to/tee_receivers_a.lst, is as follows ('id' is the pool ID and 'ip' a comma-separated list of receivers for that pool): id=1 ip=W.W.W.W:2100 id=2 ip=Y.Y.Y.Y:2100,Z.Z.Z.Z:2100 id=3 ip=[::1]:2100 ! id=1 ip=W.W.W.W:2100 tag=0 ! id=2 ip=Y.Y.Y.Y:2100,Z.Z.Z.Z:2100 tag=100 Number of tee_receivers map entries (by default 384) can be modified via maps_entries. Content can be reloaded at runtime by sending the daemon a SIGUSR2 signal (ie. "killall -USR2 nfacctd"). Selective teeing allows to filter which pool of receivers has to receive which datagrams. Tags are applied via a pre_tag_map, the one illustrated below applies tag 100 to packets exported from agents A.A.A.A, B.B.B.B and C.C.C.C; in case there was also an agent D.D.D.D exporting towards the replicator, its packets would intuitively remain untagged. Tags are matched by a tee_receivers map, see above the two pool definitions commented out containing the 'tag' keyword: the definition would cause untagged packets (tag=0) to be replicated only to pool #1 whereas packets tagged as 100 (tag=100) to be replicated only to pool #2. More examples in the pretag.map.example and tee_receivers.lst.example files in the examples/ sub-tree: set_tag=100 ip=A.A.A.A set_tag=100 ip=B.B.B.B set_tag=100 ip=C.C.C.C To enable the transparent mode, the tee_transparent should be commented out. It preserves the original IP address of the NetFlow/sFlow sender while replicating by essentially spoofing it. This feature is not global and can be freely enabled only on a subset of the active replicators. It requires super-user permissions in order to run. Concluding note: 'tee' plugin is not compatible with different plugins - within the same daemon instance. So if in the need of using pmacct for both collecting and replicating data, two separate instances must be used (intuitively with the replicator instance feeding the collector one). XVa. Splitting and dissecting data while replicating When replicating, it is possible to perform finer-grained filtering, ie. against flow-specific primitives. For example: replicate flows from or to MAC address X1, X2 .. Xn to receiver Y or replicate flows in VLAN W to receiver Z. The feature works by inspecting the original packet and dissecting it as needed, one of the most popular use-cases being IXPs replicating flows back to the members originating and/or receiving them. The feature is available for sFlow and NetFlow/IPFIX. Some of the supported primitives are: source/ destination MAC addresses and input/output interfaces ifindex; the full list of primitives is available in examples/pretag.map.example (look for "sfacctd when in 'tee' mode" or "nfacctd when in 'tee' mode"). The feature is configured just like selective teeing shown in the previous section. Incoming packets are tagged with a pre_tag_map and then matched to a receiver in tee_receivers. For example: replicate flows from/to MAC address XX:XX:XX:XX:XX:XX to receiver Y, flows from/to MAC address WW:WW:WW:WW:WW:WW to receiver Z and replicate any remaining (untagged) flows to receiver J. This is the pre_tag_map map: set_tag=100 ip=0.0.0.0/0 src_mac=XX:XX:XX:XX:XX:XX set_tag=100 ip=0.0.0.0/0 dst_mac=XX:XX:XX:XX:XX:XX set_tag=200 ip=0.0.0.0/0 src_mac=WW:WW:WW:WW:WW:WW set_tag=200 ip=0.0.0.0/0 dst_mac=WW:WW:WW:WW:WW:WW set_tag=999 ip=0.0.0.0/0 This is the tee_receivers map: id=100 ip=Y.Y.Y.Y:2100 tag=100 id=200 ip=Z.Z.Z.Z:2100 tag=200 id=999 ip=J.J.J.J:2100 tag=999 This is the relevant section from nfacctd/sfacctd configuration: [ .. ] ! tee_transparent: true maps_index: true ! plugins: tee[a] ! tee_receivers[a]: /path/to/tee_receivers.lst pre_tag_map[a]: /path/to/pretag.map Along with dissected flows, it is possible to replicate also the original packet as-is. This can be done in a separate 'tee' plugin with no or a simple pre_tag_map defined. Extending the example above, we instantiate two plugins; 'a' is as per the previous example, 'b' is to replicate full packets (as in plain, non-dissected replication): [ .. ] ! tee_transparent: true maps_index: true ! plugins: tee[a], tee[b] ! tee_receivers[a]: /path/to/tee_receivers_a.lst pre_tag_map[a]: /path/to/pretag_a.map ! tee_receivers[a]: /path/to/tee_receivers_b.lst pre_tag_map[a]: /path/to/pretag_b.map tee_receivers_a.lst and pretag_a.map are precisely as in the previous example. This is the pretag_b.map map: set_tag=100 ip=0.0.0.0/0 This is the tee_receivers_b.map map: id=100 ip=H.H.H.H:2100 tag=100 Defining a simple pre_tag_map as one where no entries require dissecting flows (ie. tagging on source_id, agent_id, ip, etc. keys) and a complex pre_tag_map as one that requires dissecting flows (ie. tagging on vlan, src_mac, dst_net, etc. keys) this is the table of truth for stopping or passing replicated data: no pre_tag_map | simple pre_tag_map | complex pre_tag_map | | dissected no | no | yes | | full yes | yes | no There are a few restrictions to the feature: 1) for sFlow, only sFlow v5 flow samples are supported, ie. no sFlow v2/v4 nor counter samples; 2) for NetFlow/ IPFIX, templates and option data is broadcasted to configured tee receivers (regardless of tags); 3) once split, flows are not muxed back together, ie. in case multiple samples part of the same packet are to be replicated to the same receiver; 4) sequence numbers are untouched: the most obvious cases being receivers may detect non-increasing sequencing or false duplicates. If you are negatively affected by any of these restrictions or limitations or you need other primitives to be supported by this feature, please get in touch. XVI. Quickstart guide to setup the BMP daemon BMP can be run as a stand-alone collector daemon (pmbmpd) or as a thread within one of the traffic accounting daemons (ie. nfacctd). The stand-alone daemon is suitable for consuming BMP data only, real-time or at regular intervals; the thread solution is suitable for correlation of BMP with other data sources, ie. NetFlow, IPFIX, sFlow, etc.. The implementation bases on rfc7854 and is in-sync with most recent RFCs and draft-ietf-grow-bmp-* IETF documents (ie. rfc8671, rfc9069, ietf-grow-bmp-tlv, ietf-grow-bmp-tlv-ebit, etc.). If unfamiliar with BMP, to quote rfc7854: "BMP is intended to provide a more convenient interface for obtaining route views for research purpose than the screen-scraping approach in common use today. The design goals are to keep BMP simple, useful, easily implemented, and minimally service-affecting.". The BMP daemon currently supports BMP data, events and stats, ie. initiation, termination, peer up, peer down, stats and route monitoring messages (that is no route mirroring messages; please get in touch if an implementation is wanted in this sense). The daemon allows to write BMP messages to files, AMQP and Kafka brokers, real-time (msglog) or at regular time intervals (dump). Also, route monitoring messages are saved in a RIB structure for IP prefix lookup. All features export data formatted as JSON or Avro messages. Please review the section "Compiling pmacct with JSON support" and, if planning to use Avro encoding, also the section "Compiling pmacct with Avro support". If writing to AMQP or Kafka brokers, compiling against RabbitMQ or Kafka libraries is required; read more in respectively the "Running the RabbitMQ/AMQP plugin" or "Running the Kafka plugin" sections of this document. If making use of the Schema Registry component of the Confluent Platform please also review the "Running the Kafka plugin" section. Following a simple example on how to configure nfacctd to enable the BMP thread to a) log, in real-time, BGP stats, events and routes received via BMP to a text-file (bmp_daemon_msglog_file) and b) dump the same (ie. BGP stats and events received via BMP) to a text-file and at regular time intervals (bmp_dump_refresh_time, bmp_dump_file): bmp_daemon: true ! bmp_daemon_msglog_file: /path/to/bmp-$peer_src_ip.log ! bmp_dump_file: /path/to/bmp-$peer_src_ip-%H%M.dump bmp_dump_refresh_time: 60 Following a simple example on how to configure nfacctd to enable the BMP thread to a) log, in real-time, BGP stats, events and routes received via BMP to a Kafka broker (bmp_daemon_msglog_kafka_topic) and b) dump the same (ie. BGP stats and events received via BMP) to a text-file and at regular time intervals (bmp_dump_refresh_time, bmp_dump_kafka_topic): bmp_daemon: true ! bmp_daemon_msglog_kafka_topic: pmacct.bmp-msglog ! bmp_daemon_msglog_kafka_broker_host: 127.0.0.1 ! bmp_dump_kafka_topic: pmacct.bmp-dump ! bmp_dump_kafka_broker_host: 127.0.0.1 bmp_dump_refresh_time: 60 ! ! [Optional] librdkafka config for buffer tuning, ssl config, etc. ! Read more in section "IXa. Notes on writing to Kafka". ! bmp_dump_kafka_config_file: /path/to/librdkafka.conf When many BMP exporters are pointed to a BMP receiver writing to Kafka, as in the example above, due to the state synchronization nature of Route Monitoring, the amount of data to be produced to the broker may exceed default queues depth. Please see the "Running the Kafka plugin" section of this document on how to adjust Kakfa queue depths via directives like batch.num.messages, queue.buffering.max.messages, etc. A sample of both the BMP msglog and dump formats are captured in the following document: docs/MSGLOG_DUMP_FORMATS Setting up the stand-alone BMP collector daemon, pmbmpd, is the exact same as the configuration above except the 'bmp_daemon: true' line can be skipped. Following is an example of how a Cisco router running IOS/IOS-XE (15.4(2)T) should be configured in order to export BMP data to a collector: router bgp 64512 bmp server 1 address X.X.X.X port-number 1790 initial-delay 60 failure-retry-delay 60 flapping-delay 60 stats-reporting-period 300 activate exit-bmp-server-mode ! neighbor Y.Y.Y.Y remote-as 64513 neighbor Y.Y.Y.Y bmp-activate all neighbor Z.Z.Z.Z remote-as 64514 neighbor Z.Z.Z.Z bmp-activate all Following is an example how a Cisco router running IOS-XR (6.2.1) should be configured in order to export BMP data to a collector: router bgp 64512 neighbor Y.Y.Y.Y bmp-activate server 1 neighbor Z.Z.Z.Z bmp-activate server 1 ! ! bmp server 1 host X.X.X.X port 1790 description -- fretta -- update-source GigabitEthernet0/0/0/10 initial-delay 60 initial-refresh delay 60 stats-reporting-period 300 initial-refresh delay 10 ! Following is an example how a Juniper router should be configured in order to export BMP data to a collector: routing-options { bmp { station FQDN { connection-mode active; monitor enable; route-monitoring { pre-policy; post-policy; } station-address X.X.X.X; station-port 1790; } } } Following is an example how a Huawei router should be configured in order to export BMP data to a collector: bmp statistics-timer 3600 # bmp-session 192.0.2.1 tcp connect port 1790 connect-interface GigabitEthernet0/3/6 # monitor peer 198.51.100.19 route-mode ipv4-family vpnv4 adj-rib-in pre-policy route-mode ipv4-family vpnv4 adj-rib-in post-policy route-mode ipv4-family vpnv4 adj-rib-out pre-policy route-mode ipv4-family vpnv4 adj-rib-out post-policy # monitor peer 198.51.100.25 route-mode ipv4-family vpnv4 adj-rib-in pre-policy route-mode ipv4-family vpnv4 adj-rib-in post-policy route-mode ipv4-family vpnv4 adj-rib-out pre-policy route-mode ipv4-family vpnv4 adj-rib-out post-policy # monitor vpn-instance A route-mode ipv4-family unicast local-rib Contribution of further configuration examples for Cisco and Juniper devices and/or other relevant vendors is more than welcome. XVII. Quickstart guide to setup Streaming Telemetry collection Quoting Cisco IOS-XR Telemetry Configuration Guide at the time of this writing: "Streaming telemetry lets users direct data to a configured receiver. This data can be used for analysis and troubleshooting purposes to maintain the health of the network. This is achieved by leveraging the capabilities of machine-to- machine communication. The data is used by development and operations (DevOps) personnel who plan to optimize networks by collecting analytics of the network in real-time, locate where problems occur, and investigate issues in a collaborative manner.". Streaming telemetry support comes in pmacct in two flavours: 1) a telemetry thread can be started in existing daemons, ie. sFlow, NetFlow/IPFIX, etc. for the purpose of data correlation and 2) a new daemon pmtelemetryd for standalone consumption of data. Streaming telemetry data can be logged real-time and/or dumped at regular time intervals to flat-files, RabbitMQ or Kafka brokers. All features export data formatted as JSON messages, hence compiling pmacct against libjansson is a requirement. See how to compile pmacct with JSON/ libjansson support in the section "Compiling pmacct with JSON support" of this document. If writing to AMQP or Kafka brokers compiling against RabbitMQ or Kafka libraries is required; read more in respectively the "Running the RabbitMQ/AMQP plugin" or "Running the Kafka plugin" sections of this document. By defining telemetry_daemon_grpc_collector_conf, Streaming Telemetry data can be optionally consumed via a thread that takes care of gRPC and GPB (Protobuf, or Google Protocol Buffers) de-marshalling and conversion to JSON; the thread leverages https://github.com/network-analytics/mdt-dialout-collector as a library; the mdt-dialout-collector GitHub repository also says how to build the stack of dependencies in order to successfully collect gRPC / GPB data: https://github.com/network-analytics/mdt-dialout-collector/blob/main/doc/integration-with-pmtelemetryd.md The gRPC thread can be configured via telemetry_daemon_grpc_collector_conf, read more in CONFIG-KEYS. From a configuration standpoint both the thread (ie. telemetry configured part of nfacctd) and the daemon (pmtelemetryd) are configured the same way except the thread must be explicitly enabled with a 'telemetry_daemon: true' config line. Hence the following examples hold for both the thread and the daemon setups. Following is a config example to receive telemetry data in JSON format over UDP port 1620 and log it real-time to flat-files: ! Telemetry thread configuration ! telemetry_daemon: true ! telemetry_daemon_port_udp: 1620 telemetry_daemon_decoder: json ! telemetry_daemon_msglog_file: /path/to/spool/telemetry-msglog-$peer_src_ip.txt ! telemetry_daemon_msglog_amqp_routing_key: telemetry-msglog ! telemetry_daemon_msglog_kafka_topic: telemetry-msglog Following is a config example to receive telemetry data with Cisco proprietary header, in GPB format over TCP port 1620 and dump it at 60 secs time intervals to flat-files: ! Telemetry thread configuration ! telemetry_daemon: true ! telemetry_daemon_port_tcp: 1620 telemetry_daemon_decoder: cisco_v1 ! telemetry_dump_file: /path/to/spool/telemetry-dump-$peer_src_ip-%Y%m%d-%H%M.txt telemetry_dump_latest_file: /path/to/spool/telemetry-dump-$peer_src_ip.latest ! telemetry_dump_amqp_routing_key: telemetry-dump ! telemetry_dump_kafka_topic: telemetry-dump ! ! [Optional] librdkafka config for buffer tuning, ssl config, etc. ! telemetry_dump_kafka_config_file: /path/to/librdkafka.conf ! telemetry_dump_refresh_time: 60 Following is a config example to receive JSON-formatted telemetry data from a pre-processor over a ZeroMQ queue (see docs/README.telemetry for further info, ie. gRPC support and GPB de-marshalling): ! Telemetry thread configuration ! telemetry_daemon: true ! telemetry_daemon_zmq_address: 127.0.0.1:50000 telemetry_daemon_decoder: json ! telemetry_daemon_msglog_file: /path/to/spool/telemetry-msglog-$peer_src_ip.txt ! telemetry_daemon_msglog_amqp_routing_key: telemetry-msglog ! telemetry_daemon_msglog_kafka_topic: telemetry-msglog ! ! [Optional] librdkafka config for buffer tuning, ssl config, etc. ! telemetry_daemon_msglog_kafka_config_file: /path/to/librdkafka.conf A sample of both the Streaming Telemetry msglog and dump formats are captured in the following document: docs/MSGLOG_DUMP_FORMATS XVIII. Running the print plugin to write to regular files and named pipes pmacct can also output to regular files and named pipes via the 'print' plugin. Dynamic filenames are supported. Output encoding is either text-based using JSON or CSV or binary schema-based using Apache Avro; encoding can be set via the 'print_output' config directive. Interval between writes can be configured via the 'print_refresh_time' config directive. An example follows on how to write to regular files in CSV format, every 15 minutes and aggregate data in 15 minutes time-bins (print_history): print_refresh_time: 900 print_history: 15m print_output: csv print_output_file: /path/to/file-%Y%m%d-%H%M.txt print_history_roundoff: m Which, over time, would produce a series of files as follows: -rw------- 1 paolo paolo 2067 Nov 21 00:15 blabla-20111121-0000.txt -rw------- 1 paolo paolo 2772 Nov 21 00:30 blabla-20111121-0015.txt -rw------- 1 paolo paolo 1916 Nov 21 00:45 blabla-20111121-0030.txt -rw------- 1 paolo paolo 2940 Nov 21 01:00 blabla-20111121-0045.txt JSON output requires compiling pmacct against Jansson library. See how to compile pmacct with JSON/libjansson support in the section "Compiling pmacct with JSON support" of this document. Avro output requires compiling pmacct against libavro library. See how to compile pmacct with Avro support in the section "Compiling pmacct with Apache Avro support" of this document. Splitting data into time-bins is supported via print_history directive. When enabled, time-related variable substitutions of dynamic print_output_file names are determined using this value. It is supported to define print_refresh_time values shorter than print_history ones by setting print_output_file_append to true (which is generally also recommended to prevent that unscheduled writes to disk, ie. due to caching issues, overwrite existing file content). A sample config follows: print_refresh_time: 300 print_output: csv print_output_file: /path/to/%Y/%Y-%m/%Y-%m-%d/file-%Y%m%d-%H%M.txt print_history: 15m print_history_roundoff: m print_output_file_append: true Writing to named pipes is an easy way to stream data into 3rd party tools without adding external dependencies (ie. RabbitMQ, Kafka). The named pipe must be manually created before starting the pmacct daemon. From a config standpoint using a named pipe is transparent and hence all the examples provided so far in this section do apply to named pipes as well. Writing to regular files is a suitable method to batch ingest data into 3rd party tools (typically, but not limited to, modern noSQL databases like ElasticSearch, InfluxDB, etc.). An example is the pmacct-to-elasticsearch project https://github.com/pierky/pmacct-to-elasticsearch . XIX. Quickstart guide to setup GeoIP lookups pmacct can perform GeoIP country lookups against a Maxmind DB v1 (--enable-geoip) and against a Maxmind DB v2 (--enable-geoipv2). A v1 database enables resolution of src_host_country and dst_host_country primitives only. A v2 database enables resolution of presently supported GeoIP-related primitives, ie. src_host_country, src_host_pocode, src_host_coords, dst_host_country, dst_host_pocode and dst_host_coords. Pre-requisite for the feature to work are: a) a working installed Maxmind GeoIP library and headers and b) a Maxmind GeoIP database (freely available). Two steps to quickly start with GeoIP lookups in pmacct: GeoIP v2 (libmaxminddb): * Have libmaxminddb library and headers to compile against, available at: https://github.com/maxmind/libmaxminddb/releases ; have also a database available: https://dev.maxmind.com/geoip/geoip2/geolite2/ . Only the database binary format is supported. * To compile the pmacct package with support for GeoIP lookups, the code must be configured for compilation as follows: shell> ./configure --enable-geoipv2 [ ... ] But, for example, should you have installed libmaxminddb in /usr/local/geoipv2 and pkg-config is unable to help, you can supply this non-default location as follows (assuming you are running the bash shell): shell> export GEOIPV2_LIBS="-L/usr/local/geoipv2/lib -lmaxminddb" shell> export GEOIPV2_CFLAGS="-I/usr/local/geoipv2/include" shell> ./configure --enable-geoipv2 [ ... ] * Include as part of the pmacct configuration the following fragment: ... geoipv2_file: /path/to/GeoIP/GeoLite2-Country.mmdb aggregate: src_host_country, dst_host_country, ... ... Concluding notes: 1) The use of --enable-geoip is mutually exclusive with --enable-geoipv2; 2) more fine-grained GeoIP lookup primitives (ie. cities, states, counties, metro areas, zip codes, etc.) are not yet supported: should you be interested into any of these, please get in touch. XX. Using pmacct as traffic/event logger This use-case has been fostered particularly by the use of NetFlow/IPFIX as generic transport, see for example Cisco NEL and Cisco NSEL. Key to logging are time-stamping primitives, timestamp_start and timestamp_end: the former records the likes of libpcap packet timestamp, sFlow sample arrival time, NetFlow observation time and flow first switched time; timestamp_end does currently only makes sense for logging flow data via NetFlow/IPFIX. Still, the exact boundary between aggregation and logging can be defined via the aggregation method, ie. no assumptions are made. An example to log traffic flows follows: ! ... ! plugins: print[traffic] ! aggregate[traffic]: src_host, dst_host, peer_src_ip, peer_dst_ip, in_iface, out_iface, timestamp_start, timestamp_end, src_port, dst_port, proto, tos, src_mask, dst_mask, src_as, dst_as, tcpflags print_output_file[traffic]: /path/to/traffic-%Y%m%d_%H%M.txt print_output[traffic]: csv print_history[traffic]: 5m print_history_roundoff[traffic]: m print_refresh_time[traffic]: 300 ! print_cache_entries[traffic]: 9999991 print_output_file_append[traffic]: true ! ! ... An example to log specifically CGNAT (Carrier Grade NAT) events from a Cisco ASR1K box follows: ! ... ! plugins: print[nat] ! aggregate[nat]: src_host, post_nat_src_host, src_port, post_nat_src_port, proto, nat_event, timestamp_start print_output_file[nat]: /path/to/nat-%Y%m%d_%H%M.txt print_output[nat]: json print_history[nat]: 5m print_history_roundoff[nat]: m print_refresh_time[nat]: 300 ! print_cache_entries[nat]: 9999991 print_output_file_append[nat]: true ! ! ... The two examples above can intuitively be merged in a single configuration so to log down in parallel both traffic flows and events. To split flows accounting from events, ie. to different files, a pre_tag_map and two print plugins can be used as follows: ! ... ! pre_tag_map: /path/to/pretag.map ! plugins: print[traffic], print[nat] ! pre_tag_filter[traffic]: 10 aggregate[traffic]: src_host, dst_host, peer_src_ip, peer_dst_ip, in_iface, out_iface, timestamp_start, timestamp_end, src_port, dst_port, proto, tos, src_mask, dst_mask, src_as, dst_as, tcpflags print_output_file[traffic]: /path/to/traffic-%Y%m%d_%H%M.txt print_output[traffic]: csv print_history[traffic]: 5m print_history_roundoff[traffic]: m print_refresh_time[traffic]: 300 ! print_cache_entries[traffic]: 9999991 print_output_file_append[traffic]: true ! pre_tag_filter[nat]: 20 aggregate[nat]: src_host, post_nat_src_host, src_port, post_nat_src_port, proto, nat_event, timestamp_start print_output_file[nat]: /path/to/nat-%Y%m%d_%H%M.txt print_output[nat]: json print_history[nat]: 5m print_history_roundoff[nat]: m print_refresh_time[nat]: 300 ! print_cache_entries[nat]: 9999991 print_output_file_append[nat]: true ! ! ... In the above configuration both plugins will log their data in 5 mins files basing on the 'print_history[]: 5m' configuration directive, ie. traffic-20130802-1345.txt traffic-20130802-1350.txt traffic-20130802-1355.txt etc. Granted appending to output file is set to true, data can be refreshed at shorter intervals than 300 secs. This is a snippet from /path/to/pretag.map referred above: set_tag=10 ip=A.A.A.A sample_type=flow set_tag=20 ip=A.A.A.A sample_type=event set_tag=10 ip=B.B.B.B sample_type=flow set_tag=20 ip=B.B.B.B sample_type=event ! ! ... XXI. Connecting pmacct to a Redis cache Since version 1.7.5, pmacct daemons can connect to a Redis cache. The main use-case currenly covered is: registering every stable daemon component (ie. plugins, threads, etc. but not writers for example) in a table so to have, when running a cluster comprising several daemons / components, an olistic view of what is currently running and where; shall a component stop running or crash it will disappear from the inventory. Over time it is expected that the amount of use-cases will increase to include specific features where one daemon can obtain relevant information about other daemons running in the cluster. Requirement to use the Redis feature plugin is the Hiredis library available here: https://github.com/redis/hiredis . Once this is installed, pmacct can be configured for compiling. pmacct makes use of pkg-config for finding libraries and headers location and checks some default locations, ie. /usr/local/lib and /usr/local/include. If this is satisfactory, all you should do is just: shell> ./configure --enable-redis But, for example, should you have installed Hiredis in /usr/local/redis and pkg-config is unable to help, you can supply this non-default location as follows (assuming you are running the bash shell): shell> export REDIS_LIBS="-L/usr/local/redis/lib -lhiredis" shell> export EDIS_CFLAGS="-I/usr/local/redis/include" shell> ./configure --enable-redis Following a configuration snippet showing a basic configuration for the Redis feature. redis_host points to the Redis cache whereas cluster_* directives essentially form a pmacct cluster (cluster_name, every cluster should have a different name string, and cluster_id, an unique id inside a certain cluster): cluster_name: prod cluster_id: 1 ! redis_host: 127.0.0.1:6379 XXII. Miscellaneous notes and troubleshooting tips This chapter will hopefully build up to the point of providing a taxonomy of popular cases to troubleshoot by daemon and what to do. Although that is the plan, the current format is sparse notes. When reporting a bug: please report in all cases the pmacct version that you are experiencing your issue against; the CLI option -V of the daemon you are using returns all the info needed (daemon, version, specific release and options compiled in). Do realise that if using a pre-packaged version from your OS and/or old code (ie. not master code on GitHub or latest official release), you may be very possibly asked to try one of these first. Finally, please refrain to open issues on GitHub if not using master code (use the pmacct-discussion mailing list or unicast email instead). a) In cases of errors at configure script time (ie. library not found, headers not found, etc.), additional info can be found in config.log at the end of the 'Core tests' section. Also if reporting an issue related to configure time, it is good idea to supply the config.log file as well. Here are recap some popular issues when compiling pmacct or linking it at runtime against shared libraries: 1) /usr/local/sbin/pmacctd: error while loading shared libraries: librabbitmq.so.4: cannot open shared object file: No such file or directory This can happen at runtime and, especially in case of freshly downloaded and compiled libraries, it is a symptom that after installing the shared library, ldconfig was not called. Or alternatively that the directory where the library is located is not inserted in /etc/ld.so.conf or in any files included it includes. When not having super-user privileges, this can be fixed in the own environment setting the LD_LIBRARY_PATH variable. 2) json_array_foreach(json_list, key, value) { ^ nfv9_template.c: In function ‘nfacctd_offline_read_json_template’: nfv9_template.c:572:53: error: expected ‘;’ before ‘{’ token This can happen at compile time and is a bit tricky to hint. In this example the function json_array_foreach() is not being recognized, in other words while the library could be located, it does not contain the specific function. This is a symptom that the library version in use is too old. Typical situation is when using a packaged library rather than a freshly downloaded and compiled latest stable release. 3) /usr/local/lib/libpcap.so: undefined reference to `pcap_lex' collect2: error: ld returned 1 exit status make[2]: *** [pmacctd] Error 1 This can happen at compile time and it is a symptom that the needed library could not be located by the linker. This is a symptom that the library could be in some non-standard location and the linked need an hint. For libpcap --with-pcap-libs knob is available at configure time; for all other libraries the library_LIBS and library_CFLAGS environment variables are available. See examples in the "Configuring pmacct for compilation and installing" section of this document. b) In case of crashes of an any process, regardless if predictable or not, the advice is to run the daemon with "ulimit -c unlimited" so to generate a core dump. The core dump file location, on Linux, is determined by the following file: /proc/sys/kernel/core_pattern : if the value is "core", then the file is placed in the directory where the daemon is started so it is good to take care of that. pmacct developers will then ask for one or both of the following: 1) the core file along with the crashing executable and its configuration be made available for further inspection and/or 2) a backtrace in GDB obtained via the following two steps: shell> gdb /path/to/executable /path/to/core Then once in the gdb console the backtrace output can be obtained with the following command: gdb> bt As an alternative to gdb, the valgrind tool can be used which, upon crashing, will simply print out the backtrace: shell> valgrind /path/to/executable If the dmesg command or system logs reveal a line like the following one: Jan 01 00:00:00 xxxx kernel: [4075819.569587] nfacctd[51817]: segfault at 10 ip 00007f77f448ad44 sp 00007ffc6974ed70 error 4 .. In certain cases, ie. if the crash is not inside a shared library, this can return useful information for debugging when examined with addr2line; using the example below: shell> addr2line /path/to/nfacctd 00007f77f448ad44 Optionally, especially if the issue can be easily reproduced, the daemon can be re-configured for compiling with the --enable-debug flag so to produce extra info suitable for troubleshooting. There is a whole range of cases where the crash could be caused by corrupting the heap and these are evidenced when a 'bt' in gdb highlights a malloc() call as point of crash; while these issues often times boil down to erroneous coding, they are subtle, hard to reproduce and platform / OS dependant. When facing such issues, launching the daemon with the MALLOC_CHECK_ environment variable set to 2, ie.: shell> MALLOC_CHECK_=2 shell> export MALLOC_CHECK_ Would return additional useful information for the troubleshooting process. At times this action - that elicits the use of an alternate malloc() call - may even work as a work-around (if falling in this condition it is recommended to not use this 'mask' as long-term solution, that is, please report the issue). The use of Docker containers is getting increasingly popular. The valgrind tool can be used when a crash is experienced in a container. The following example applies to any of the official pmacct containers (running on top of Debian): shell> docker run -it --entrypoint bash # apt-get update && apt-get install -y valgrind # valgrind /path/to/executable c) In case of (suspected) memory leaks, the advice is to: 1) re-compile pmacct with "./configure --enable-debug "; --enable-debug sets as CFLAGS -O0 -g -Wall where especially -O0 is capital since it disables any code optimizations the compiler may introduce; 2) run the resulting daemon under valgrind, "valgrind --leak-check=yes /path/to/executable ". A memory leak is confirmed if the amount of "definitely lost" bytes keeps increasing over time. d) In the two cases of nfacctd/sfacctd or nfprobe/sfprobe not showing signs of input/output data: 1) check with tcpdump, ie. "tcpdump -i -n " (for uacctd use -i nflog:), that packets are emitted/received. Optionally Wireshark (or its cli tool tshark) can be used, in conjunction with decoders ('cflow' for NetFlow/ IPFIX and 'sflow' for sFlow), to validate packets are consistent; this proofs there is no filtering taking place in between exporters and collector; a sample command-line for tshark for NetFlow/IPFIX would be: tshark -d udp.port==,cflow -VVV src host Optionally -Y can be passed to further filter flow traffic down and/or validate specific elements of it, ie. -Y "cflow.dstaddr==X.X.X.X/31". 2) check firewall settings on the collector box, ie. "iptables -L -n" on Linux (disable or do appropriate holes): tcpdump may see packets hitting the listening port as, in normal kernel operations, the filtering happens after the raw socket (the one used by tcpdump) is served; you can also check with 3rd party equivalent tools or netcat, "nc -l -u -p ", that the same behaviour is obtained as with pmacct; finally, if running on a SELinux enabled system, consider that it could be interfering too; 3) check the process list to confirm whether there is some stuck / forgotten running process which may be consuming data in place of a newly started one; 4) when not using ZeroMQ for internal buffering (plugin_pipe_zmq set to true) and especially in case of copy/paste of configs or if using a config from a production system in lab, disable/double-check values for internal buffering: if set too high they will likely retain data internally to the daemon; 5) if multiple interfaces are configured on a system, try to disable rp_filter. See http://tldp.org/HOWTO/Adv-Routing-HOWTO/lartc.kernel.rpf.html for more info on RP filtering. The value of rp_filter files in /proc must be set to zero (ie. "echo 0 > /proc/sys/net/ipv4/conf/all/rp_filter") in order to disable RP filtering. Also, if working on a cloud instance, consider often times these are not configured by default to pass traffic through. These appear popular issues when replicating flow data with tee_transparent feature enabled (that is, packet spoofing); 6) in case aggregate_filter is in use: the feature expects a libpcap-style filter as value. BPF filters are sensible to both VLAN tags and MPLS labels: if, for example, the traffic is VLAN tagged and the value of aggregate_filter is 'src net X.X.X.X/Y', there will be no match for VLAN-tagged traffic from src net X.X.X.X/Y; the filter should be re-written as 'vlan and src net X.X.X.X/Y'; 7) in case of NetFlow v9/IPFIX collection, two protocols that are template- based, the issue may be with templates not being received by nfacctd (in which case by enabling debug you may see "Discarded NetFlow v9/IPFIX packet (R: unknown template [ .. ]" messages in your logs); you can confirm whether templates are being exported/replicated/received with: * tshark / Wireshark, ie. "tshark -d udp.port==,cflow -R cflow.template_id -2"; * nfacctd, either enabling debug - in which case templates will be part of the log output - or using nfacctd_templates_file - which will dump templates to the pointed file in JSON format; it may also happen the opposite: templates are received but no data packets are being sent (ie. because sampling is not configured on the actual interfaces on the exporting device): in this case tshark would only show periodic data being received (ie. every 30 secs, 60 secs, etc.) rather than a more constant stream of data. e) In the case of inconsistent output - inconsistent defined as replaying the same set of data, output produced by the daemon differs - of nfacctd/ sfacctd: 1) check the basics: whether another nfacctd/sfacctd is listening on the very same port; this is legit behaviour due to behaviour of SO_REUSEPORT, in brief: share load among multiple sockets listening on the same port, and SO_REUSEADDR, in brief: don't wait for a closing socket to be closed in order to bind a new one to the same port, options used on the listening socket; 2) check all tips about buffering provided slightly later in this document. f) Replay packets can be needed, for example, to troubleshoot the behaviour of one of the pmacct daemons. A capture in libpcap format, suitable for replay, can be produced with tcpdump. For example to produce a trace for NetFlow/IPFIX/sFlow this is the command-line: "tcpdump -i -n -s 0 -w port "; for ULOG/ NFLOG on Linux the interface should be in the format "-i nflog:" instead. For traces over extended periods of time, ie. to troubleshoot specific bogus packets triggering crashes or capture NetFlow v9/IPFIX templates, optionally the -G tcpdump option can be specified to rotate the -w output_file each specified seconds interval (read more on tcpdump man page). Once the trace has been created, trace file(s) can be replayed by pmacct daemons using the pcap_savefile (-I) and, optionally, the pcap_savefile_wait (-W) directives, ie.: "nfacctd -I <..>". For more advanced use-cases, ie. loop indefinitely through the pcap file and run it with a speed multiplicator in order to stress test the daemon, the tcpreplay tool can be used. In this case, before replaying NetFlow/ IPFIX/sFlow, L2/L3 of captured packets may need to be adjusted in order to reflect the lab/replay environment; this can be done with the tcprewrite tool of the tcpreplay package, ie.: "tcprewrite --enet-smac= --enet-dmac= -S -D --fixcsum --infile= --outfile=". Then the output file from tcprewrite can be supplied to tcpreplay for the actual replay to the pmacct daemon, ie.: "tcpreplay -x -i ". g) Buffering is often an element to tune in order to make sure that output is complete (packet loss may lead to incomplete results but also configuring too large buffers). While buffering internal to pmacct (configured with plugin_buffer_size / plugin_pipe_size or plugin_pipe_zmq) returns warning messages in case of data loss, buffering between pmacct and the kernel (configured with nfacctd_pipe_size and its equivalents) may be more tricky to detect and sometimes issues can only be inferred by symptoms (like sequence number checks for NetFlow v9/IPFIX). For nfacctd and sfacctd daemons, two commands useful to check this kind of buffering on Linux systems are: 1) "cat /proc/net/udp" or "cat /proc/net/udp6" ensuring that "drops" value is not increasing and 2) "netstat -s" ensuring, under the section UDP, that errors are not increasing (since this command returns system-wide counters, the counter-check would be: stop the pmacct daemon running and, granted the counter was increasing, verify it does not increase anymore). For pmacctd, an UNIX signal can be sent to the daemon so that the amount of dropped packets is reported in the log output (ie. kill -USR1 ). Finally, if troubleshooting incomplete output, please review tips about inconsistent output slightly earlier in this document. h) BMP, the BGP Monitoring Protocol, has got drafting traction at IETF. New protocol features drive the need for inter-op testing. The BMP play tool ( https://github.com/paololucente/bmp_play ) is a simple script to record and replay raw BMP packets in order to reproduce issues. An example of how this can be done, here: Step 1: export BMP data from an exporter to a host/port where a bmp_play script is listening to. bmp_play will then save raw BMP data to a file: shell> bmp_play.py -m record -p 1791 -f bmp.dump Step 2: the file containing raw BMP data can be replayed at will with the same bmp_play script to a pmbmpd daemon: shell 1> bmp_play.py -m play -p 1790 -f bmp.dump -d 127.0.0.1 shell 2> pmbmpd -l 1790 -d -o bmp.log Additionally pmbmpd supports pcap_savefile (-I) to replay a pcap trace that includes BMP packets; packets functional to the house- keeping of the TCP session are safely ignored and pcap_filter helps on replay to trash irrelevant (non BMP) packets; yet this method is less preferred to the capture of raw data via bmp_play. i) if supported and if enabled on the system, the SO_REUSEPORT feature can be leveraged on Linux and most BSDs where it is part of the standard kernel: it allows multiple daemons to bind to the same local address and port in order to load-balance processing of incoming packets. Here a reference to a presentation of this feature for Linux: https://domsch.com/linux/lpc2010/Scaling_techniques_for_servers_with_high_connection%20rates.pdf To enable SO_REUSEPORT on a Linux system 'sysctl net.core.allow_reuseport=1'. Also the processes must be launched by the same user.