fastnetmon-ng/docs/FINE_TUNING.md
2015-06-16 17:24:13 +03:00

3.7 KiB

I recommend you to disable CPU freq scaling for gain max performance (max frequency):

echo performance | tee /sys/devices/system/cpu/cpu*/cpufreq/scaling_governor

You can use script irq_balance_manually.sh for irq balancing on heavy loaded networks.

Running tool without root permissions:

useradd fastnetmon
setcap cap_net_admin+eip fastnetmon
su fastnetmon
./fastnetmon eth0,eth1

Please keep in mind when run tool on OpenVZ because without root permissions tool can't get all VE ips and you should pass it explicitly.

Debugging flags.

DUMP_ALL_PACKETS will enable all packets dumping to console. It's very useful for testing tool on non standard platforms.

DUMP_ALL_PACKETS=yes ./fastnetmon

Recommended configuration options for ixgbe Intel X540 driver (netmap mode):

cat /etc/modprobe.d/ixgbe.conf
options ixgbe IntMode=2,2 MQ=1,1 DCA=2,2 RSS=8,8 VMDQ=0,0 max_vfs=0,0 L2LBen=0,0 InterruptThrottleRate=1,1 FCoE=0,0 LRO=1,1 allow_unsupported_sfp=0,0

I got very big packet size (more than mtu) in attack log? In PF_RING this behaviour will be related with offload features of NIC. For INtel 82599 I recommend disable all offload:

ethtool -K eth0 gro off gso off tso off

How I can enable hardware filtration for Intel 82599 NIC? Install patched ixgbe driver from PF_RING distro and apply this patch to CMakeLists.txt and recompile tool:

add_definitions(-DHWFILTER_LOCKING)

How I can compile FastNetMon without PF_RING support?

cmake .. -DDISABLE_PF_RING_SUPPORT=ON

If you saw intel_idle in perf top with red higlihting you can disable it with following kernel params (more details you can find Performance_Tuning_Guide_for_Mellanox_Network_Adapters.pdf):

intel_idle.max_cstate=0 processor.max_cstate=1

If you want build with clang:

cmake -DCMAKE_C_COMPILER=/usr/bin/clang -DCMAKE_CXX_COMPILER=/usr/bin/clang++ ..

If tou want build tool with debug info:

cmake -DCMAKE_BUILD_TYPE=Debug  ..

If you want speedup build process please build with ninja instead of make:

apt-get install -y ninja-build
cd build
cmake -GNinja ..
ninja

Ninja use all CPUs for build process:

1  [||||||||||||||||||||||||||||||||||||||||||||||100.0%]     Tasks: 53, 103 thr, 64 kthr; 6 running
2  [||||||||||||||||||||||||||||||||||||||||||||||100.0%]     Load average: 1.32 0.45 0.19 
3  [||||||||||||||||||||||||||||||||||||||||||||||100.0%]     Uptime: 1 day, 12:58:40
4  [||||||||||||||||||||||||||||||||||||||||||||||100.0%]

Build script for reading Netflow (v5, v9, ipfix) data from pcap dump:

cmake .. -DBUILD_PCAP_READER=ON

Run pcap data:

./fastnetmon_pcap_reader dump.pcap

Build script for running packet capture plugins without analyzer backend:

cmake .. -DBUILD_PLUGIN_RUNNER=ON

Examples for different plugins (plugin name could be netflow, netmap, sflow, pfring, pcap):

./fastnetmon_plugin_runner netflow

How to collect data for debugging netflow:

tcpdump -w /root/netflow_data.pcap -n 'udp dst port 2055' 

How to collect data for debugging sFLOW:

 tcpdump -w /root/sflow_data.pcap -n 'udp dst port 6343'

Performance tuning:

  • Do not use short prefixes (lesser then /24)
  • Do not use extremely big prefixes (/8, /16) because memory consumption will be very big

How I can enable ZC support for PF_RING? Please install DNA/ZC dreivers, load they and add interface name with zc prefix in config file (i.e. zc:eth3)

For development new code, please check .clang-format as code guide example.

You can find more info and graphics here