doc/userguide/performance/high-performance-config.rst
One of the major dependencies for Suricata's performance is the Network Interface Card. There are many vendors and possibilities. Some NICs have and require their own specific instructions and tools of how to set up the NIC. This ensures the greatest benefit when running Suricata. Vendors like Napatech, Netronome, Accolade, Myricom include those tools and documentation as part of their sources.
For Intel, Mellanox and commodity NICs the following suggestions below could be utilized.
It is recommended that the latest available stable NIC drivers are used. In
general when changing the NIC settings it is advisable to use the latest
ethtool version. Some NICs ship with their own ethtool that is
recommended to be used. Here is an example of how to set up the ethtool
if needed:
::
wget https://mirrors.edge.kernel.org/pub/software/network/ethtool/ethtool-5.2.tar.xz tar -xf ethtool-5.2.tar.xz cd ethtool-5.2 ./configure && make clean && make && make install /usr/local/sbin/ethtool --version
When doing high performance optimisation make sure irqbalance is off and
not running:
::
service irqbalance stop
Depending on the NIC's available queues (for example Intel's x710/i40 has 64 available per port/interface) the worker threads can be set up accordingly. Usually the available queues can be seen by running:
::
/usr/local/sbin/ethtool -l eth1
Some NICs - generally lower end 1Gbps - do not support symmetric hashing see
:doc:packet-capture. On those systems due to considerations for out of order
packets the following setup with af-packet is suggested (the example below
uses eth1):
::
/usr/local/sbin/ethtool -L eth1 combined 1
then set up af-packet with number of desired workers threads threads: auto
(auto by default will use number of CPUs available) and
cluster-type: cluster_flow (also the default setting)
For higher end systems/NICs a better and more performant solution could be
utilizing the NIC itself a bit more. x710/i40 and similar Intel NICs or
Mellanox MT27800 Family [ConnectX-5] for example can easily be set up to do
a bigger chunk of the work using more RSS queues and symmetric hashing in order
to allow for increased performance on the Suricata side by using af-packet
with cluster-type: cluster_qm mode. In that mode with af-packet all packets
linked by network card to a RSS queue are sent to the same socket. Below is
an example of a suggested config set up based on a 16 core one CPU/NUMA node
socket system using x710:
::
rmmod i40e && modprobe i40e ifconfig eth1 down /usr/local/sbin/ethtool -L eth1 combined 16 /usr/local/sbin/ethtool -K eth1 rxhash on /usr/local/sbin/ethtool -K eth1 ntuple on ifconfig eth1 up /usr/local/sbin/ethtool -X eth1 hkey 6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A equal 16 /usr/local/sbin/ethtool -A eth1 rx off /usr/local/sbin/ethtool -C eth1 adaptive-rx off adaptive-tx off rx-usecs 125 /usr/local/sbin/ethtool -G eth1 rx 1024
The commands above can be reviewed in detail in the help or manpages of the
ethtool. In brief the sequence makes sure the NIC is reset, the number of
RSS queues is set to 16, load balancing is enabled for the NIC, a low entropy
toeplitz key is inserted to allow for symmetric hashing, receive offloading is
disabled, the adaptive control is disabled for lowest possible latency and
last but not least, the ring rx descriptor size is set to 1024.
Make sure the RSS hash function is Toeplitz:
::
/usr/local/sbin/ethtool -X eth1 hfunc toeplitz
Let the NIC balance as much as possible:
::
for proto in tcp4 udp4 tcp6 udp6; do /usr/local/sbin/ethtool -N eth1 rx-flow-hash $proto sdfn done
In some cases:
::
/usr/local/sbin/ethtool -N eth1 rx-flow-hash $proto sd
might be enough or even better depending on the type of traffic. However not
all NICs allow it. The sd specifies the multi queue hashing algorithm of
the NIC (for the particular proto) to use src IP, dst IP only. The sdfn
allows for the tuple src IP, dst IP, src port, dst port to be used for the
hashing algorithm.
In the af-packet section of suricata.yaml:
::
af-packet:
Intel based systems
If the system has more then one NUMA node there are some more possibilities.
In those cases it is generally recommended to use as many worker threads as
cpu cores available/possible - from the same NUMA node. The example below uses
a 72 core machine and the sniffing NIC that Suricata uses located on NUMA node 1.
In such 2 socket configurations it is recommended to have Suricata and the
sniffing NIC to be running and residing on the second NUMA node as by default
CPU 0 is widely used by many services in Linux. In a case where this is not
possible it is recommended that (via the cpu affinity config section in
suricata.yaml and the irq affinity script for the NIC) CPU 0 is never used.
In the case below 36 worker threads are used out of NUMA node 1's CPU,
af-packet runmode with ``cluster-type: cluster_qm``.
If the CPU's NUMA set up is as follows:
::
lscpu
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
CPU(s): 72
On-line CPU(s) list: 0-71
Thread(s) per core: 2
Core(s) per socket: 18
Socket(s): 2
NUMA node(s): 2
Vendor ID: GenuineIntel
CPU family: 6
Model: 79
Model name: Intel(R) Xeon(R) CPU E5-2697 v4 @ 2.30GHz
Stepping: 1
CPU MHz: 1199.724
CPU max MHz: 3600.0000
CPU min MHz: 1200.0000
BogoMIPS: 4589.92
Virtualization: VT-x
L1d cache: 32K
L1i cache: 32K
L2 cache: 256K
L3 cache: 46080K
NUMA node0 CPU(s): 0-17,36-53
NUMA node1 CPU(s): 18-35,54-71
It is recommended that 36 worker threads are used and the NIC set up could be
as follows:
::
rmmod i40e && modprobe i40e
ifconfig eth1 down
/usr/local/sbin/ethtool -L eth1 combined 36
/usr/local/sbin/ethtool -K eth1 rxhash on
/usr/local/sbin/ethtool -K eth1 ntuple on
ifconfig eth1 up
./set_irq_affinity local eth1
/usr/local/sbin/ethtool -X eth1 hkey 6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A equal 36
/usr/local/sbin/ethtool -A eth1 rx off tx off
/usr/local/sbin/ethtool -C eth1 adaptive-rx off adaptive-tx off rx-usecs 125
/usr/local/sbin/ethtool -G eth1 rx 1024
for proto in tcp4 udp4 tcp6 udp6; do
echo "/usr/local/sbin/ethtool -N eth1 rx-flow-hash $proto sdfn"
/usr/local/sbin/ethtool -N eth1 rx-flow-hash $proto sdfn
done
In the example above the ``set_irq_affinity`` script is used from the NIC
driver's sources.
In the cpu affinity section of suricata.yaml config:
::
# Suricata is multi-threaded. Here the threading can be influenced.
threading:
cpu-affinity:
management-cpu-set:
cpu: [ "1-10" ] # include only these CPUs in affinity settings
receive-cpu-set:
cpu: [ "0-10" ] # include only these CPUs in affinity settings
worker-cpu-set:
cpu: [ "18-35", "54-71" ]
mode: "exclusive"
prio:
low: [ 0 ]
medium: [ "1" ]
high: [ "18-35","54-71" ]
default: "high"
In the af-packet section of suricata.yaml config :
::
- interface: eth1
# Number of receive threads. "auto" uses the number of cores
threads: 18
cluster-id: 99
cluster-type: cluster_qm
defrag: no
mmap-locked: yes
tpacket-v3: yes
ring-size: 100000
block-size: 1048576
- interface: eth1
# Number of receive threads. "auto" uses the number of cores
threads: 18
cluster-id: 99
cluster-type: cluster_qm
defrag: no
mmap-locked: yes
tpacket-v3: yes
ring-size: 100000
block-size: 1048576
That way 36 worker threads can be mapped (18 per each af-packet interface slot)
in total per CPUs NUMA 1 range - 18-35,54-71. That part is done via the
``worker-cpu-set`` affinity settings. ``ring-size`` and ``block-size`` in the
config section above are decent default values to start with. Those can be
better adjusted if needed as explained in :doc:`tuning-considerations`.
AMD based systems
~~~~~~~~~~~~~~~~~
Another example can be using an AMD based system where the architecture and
design of the system itself plus the NUMA node's interaction is different as
it is based on the HyperTransport (HT) technology. In that case per NUMA
thread/lock would not be needed. The example below shows a suggestion for such
a configuration utilising af-packet, ``cluster-type: cluster_flow``. The
Mellanox NIC is located on NUMA 0.
The CPU set up is as follows:
::
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
CPU(s): 128
On-line CPU(s) list: 0-127
Thread(s) per core: 2
Core(s) per socket: 32
Socket(s): 2
NUMA node(s): 8
Vendor ID: AuthenticAMD
CPU family: 23
Model: 1
Model name: AMD EPYC 7601 32-Core Processor
Stepping: 2
CPU MHz: 1200.000
CPU max MHz: 2200.0000
CPU min MHz: 1200.0000
BogoMIPS: 4391.55
Virtualization: AMD-V
L1d cache: 32K
L1i cache: 64K
L2 cache: 512K
L3 cache: 8192K
NUMA node0 CPU(s): 0-7,64-71
NUMA node1 CPU(s): 8-15,72-79
NUMA node2 CPU(s): 16-23,80-87
NUMA node3 CPU(s): 24-31,88-95
NUMA node4 CPU(s): 32-39,96-103
NUMA node5 CPU(s): 40-47,104-111
NUMA node6 CPU(s): 48-55,112-119
NUMA node7 CPU(s): 56-63,120-127
The ``ethtool``, ``show_irq_affinity.sh`` and ``set_irq_affinity_cpulist.sh``
tools are provided from the official driver sources.
Set up the NIC, including offloading and load balancing:
::
ifconfig eno6 down
/opt/mellanox/ethtool/sbin/ethtool -L eno6 combined 15
/opt/mellanox/ethtool/sbin/ethtool -K eno6 rxhash on
/opt/mellanox/ethtool/sbin/ethtool -K eno6 ntuple on
ifconfig eno6 up
/sbin/set_irq_affinity_cpulist.sh 1-7,64-71 eno6
/opt/mellanox/ethtool/sbin/ethtool -X eno6 hfunc toeplitz
/opt/mellanox/ethtool/sbin/ethtool -X eno6 hkey 6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A:6D:5A
In the example above (1-7,64-71 for the irq affinity) CPU 0 is skipped as it is usually used by default on Linux systems by many applications/tools.
Let the NIC balance as much as possible:
::
for proto in tcp4 udp4 tcp6 udp6; do
/usr/local/sbin/ethtool -N eth1 rx-flow-hash $proto sdfn
done
In the cpu affinity section of suricata.yaml config :
::
# Suricata is multi-threaded. Here the threading can be influenced.
threading:
set-cpu-affinity: yes
cpu-affinity:
management-cpu-set:
cpu: [ "120-127" ] # include only these cpus in affinity settings
receive-cpu-set:
cpu: [ 0 ] # include only these cpus in affinity settings
worker-cpu-set:
cpu: [ "8-55" ]
mode: "exclusive"
prio:
high: [ "8-55" ]
default: "high"
In the af-packet section of suricata.yaml config:
::
- interface: eth1
# Number of receive threads. "auto" uses the number of cores
threads: 48 # 48 worker threads on cpus "8-55" above
cluster-id: 99
cluster-type: cluster_flow
defrag: no
mmap-locked: yes
tpacket-v3: yes
ring-size: 100000
block-size: 1048576
In the example above there are 15 RSS queues pinned to cores 1-7,64-71 on NUMA
node 0 and 40 worker threads using other CPUs on different NUMA nodes. The
reason why CPU 0 is skipped in this set up is as in Linux systems it is very
common for CPU 0 to be used by default by many tools/services. The NIC itself in
this config is positioned on NUMA 0 so starting with 15 RSS queues on that
NUMA node and keeping those off for other tools in the system could offer the
best advantage.
.. note:: Performance and optimization of the whole system can be affected upon regular NIC driver and pkg/kernel upgrades so it should be monitored regularly and tested out in QA/test environments first. As a general suggestion it is always recommended to run the latest stable firmware and drivers as instructed and provided by the particular NIC vendor.
Other considerations
Another advanced option to consider is the isolcpus kernel boot parameter
is a way of allowing CPU cores to be isolated for use of general system
processes. That way ensures total dedication of those CPUs/ranges for the
Suricata process only.
stream.wrong_thread / tcp.pkt_on_wrong_thread are counters available
in stats.log or eve.json as event_type: stats that indicate issues with
the load balancing. There could be traffic/NICs settings related as well. In
very high/heavily increasing counter values it is recommended to experiment
with a different load balancing method either via the NIC or for example using
XDP/eBPF. There is an issue open
https://redmine.openinfosecfoundation.org/issues/2725 that is a placeholder
for feedback and findings.