Linux. Network Tuning

The buffer size options should be defined in the /etc/sysctl.conf file.
Recommended to use next values for 1G ethernet adapters:

net.core.rmem_max = 16777216
net.core.wmem_max = 16777216
net.ipv4.udp_mem = 8388608 12582912 16777216
net.ipv4.tcp_rmem = 4096 87380 8388608
net.ipv4.tcp_wmem = 4096 65536 8388608
net.core.wmem_default = 16777216
net.core.rmem_default = 16777216
net.ipv4.tcp_tw_reuse = 1

For 10G ethernet adapters:

net.core.rmem_max = 67108864
net.core.wmem_max = 67108864
net.ipv4.udp_mem = 8388608 16777216 33554432
net.ipv4.tcp_rmem = 4096 87380 33554432
net.ipv4.tcp_wmem = 4096 65536 33554432
net.core.wmem_default = 33554432
net.core.rmem_default = 33554432
net.ipv4.tcp_congestion_control=htcp
net.ipv4.tcp_tw_reuse = 1

For 40G ethernet adapters:

net.core.rmem_max = 134217728
net.core.wmem_max = 134217728
net.ipv4.udp_mem = 8388608 33554432 67108864
net.ipv4.tcp_rmem = 4096 87380 67108864
net.ipv4.tcp_wmem = 4096 65536 67108864
net.core.wmem_default = 67108864
net.core.rmem_default = 67108864
net.ipv4.tcp_congestion_control=htcp
net.ipv4.tcp_tw_reuse = 1

To apply changes restart system or launch:

sysctl -p

You can verify the current values with next command:

sysctl \
    net.core.rmem_default \
    net.core.rmem_max \
    net.core.wmem_default \
    net.core.wmem_max \
    net.ipv4.udp_mem \
    net.ipv4.tcp_wmem

[root@astra ~]# ethtool -g eth1
Ring parameters for eth1:
Pre-set maximums:
RX:     4096
RX Mini:    0
RX Jumbo:   0
TX:     4096
Current hardware settings:
RX:     4096
RX Mini:    0
RX Jumbo:   0
TX:     256

Here we can see the rx-buffer increased by the maximum. Usually it is quite difficult to find the correct value. The most optimal is average value. With a high-frequency and multi-core processor (>3GHz), you can get good results with maximum buffer size. An example command to increase the buffer:

ethtool -G eth1 rx 2048

Add to the /etc/sysctl.conf file next lines:

net.ipv4.conf.eth0.rp_filter = 2

Where eth0 is the interface name.

To apply changes restart system or launch:

sysctl -p

The IGMP version could be defined in the /etc/sysctl.conf file. For example setup IGMPv2 for eth1 interface:

net.ipv4.conf.eth1.force_igmp_version=2

To apply changes restart system or launch:

sysctl -p

You can verify IGMP version with tcpdump. Launch:

tcpdump -i eth1 igmp

Then try to subscribe to the multicast stream. For example:

astra --analyze udp://eth1@239.255.1.1:1234

Keywords: missed, dropped, fifo, error, rx.

ip -s -s link show eth1

You need to look at RX Errors. Some network cards provide more detailed information about the nature of the loss:

ethtool -S eth1

Losses can be not only on the network cards of your server. They can also be on the network equipment port. You can find the informaton how to see it in the documentation of the network equipment manufacturer.

Install

yum install python3
pip3 install netutils-linux

network-top

This utility is needed to evaluate the applied settings and displays the uniformity of the load distribution (interrupts, softirqs, the number of packets per second per processor core) on the server resources, all kinds of packet processing errors. Values that exceed the thresholds are highlighted.

rss-ladder

# rss-ladder eth1 0
- distributing interrupts of eth1 (-TxRx) on socket 0:"
  - eth1: irq 67 eth1-TxRx-0 -> 0
  - eth1: irq 68 eth1-TxRx-1 -> 1
  - eth1: irq 69 eth1-TxRx-2 -> 2
  - eth1: irq 70 eth1-TxRx-3 -> 3
  - eth1: irq 71 eth1-TxRx-4 -> 8
  - eth1: irq 72 eth1-TxRx-5 -> 9
  - eth1: irq 73 eth1-TxRx-6 -> 10
  - eth1: irq 74 eth1-TxRx-7 -> 11

This utility distributes network card interrupts to the cores of the selected physical processor (default is 0).

server-info

# server-info --rate
cpu:
  BogoMIPS: 7
  CPU MHz: 7
  CPU(s): 1
  Core(s) per socket: 1
  L3 cache: 1
  Socket(s): 10
  Thread(s) per core: 10
  Vendor ID: 10
 disk:
   vda:
     size: 1
     type: 1
 memory:
   MemTotal: 1
   SwapTotal: 10
 net:
   eth1:
     buffers:
       cur: 5
       max: 10
     driver: 1
     queues: 1
 system:
   Hypervisor vendor: 1
   Virtualization type: 1

This utility allows you to do two things:

server-info --show: see what hardware is installed on the server. In General, it is similar to lshw, but with an emphasis on the parameters of interest to us.

server-info --rate: find bottlenecks in server hardware. In General, it is similar to the Windows performance index, but with an emphasis on the parameters of interest to us. The assessment is made on a scale from 1 to 10.

Other utilities

rx-buffers-increase eth1

automatically increases the buffer of the selected network card to the optimal value.

maximize-cpu-freq

disables the floating frequency of the processor.

Example 1. As simple as possible

Task:

• Single processor with 4 cores.
• Single 1 Gbps network card (eth0) with 4 combined queues
• incoming traffic 600 Mbit/s, no outgoing traffic
• all queues hang on CPU0, a total of 55,000 interrupts and 350,000 packets per second, of which about 200 packets/sec are lost by the network card. The remaining 3 cores are idle

Decision:

distribute the queues between the cores with the command rss-ladder eth0
increase buffer with command rx-buffers-increase eth0

Example 2

Task:

• Two processors with 8 cores
• Two NUMA-nodes
• Two dual-port 10 Gbps network cards (eth0, eth1, eth2, eth3), each port has 16 queues, all tied to node 0, incoming traffic volume: 3 Gbps per each
• Single 1 Gbps network card, 4 queues, tied to node 0, outgoing traffic: 100 Mbps.

Decision:

1. put one of the 10 Gbit/s network cards to another PCI slot, bound to NUMA node 1.
2. Reduce the number of combined queues for 10 Gigabit ports to the number of cores per physical processor:

for dev in eth0 eth1 eth2 eth3; do
  ethtool -L $dev combined 8
done

3. Distribute interrupts of ports eth0, eth1 on the processor cores getting to NUMA node 0, and ports eth2, eth3 on the processor cores getting to NUMA node1:

rss-ladder eth0 0
rss-ladder eth1 0
rss-ladder eth2 1
rss-ladder eth3 1

4. Increase eth0, eth1, eth2, eth3 RX buffers:

for dev in eth0 eth1 eth2 eth3; do
  rx-buffers-increase $dev
done

Reminder:

In the case of network cards with a single queue, you can use RPS to distribute the load between the cores, but this does not eliminate the loss of copying packets into memory.

The distribution of interrupts is based on the calculation of the hash function (the remainder of the division) from the set of protocol, source and destination IP, and source and destination port. The technology is called: Receive-side scaling (RSS).