Skip to content

Latest commit

 

History

History
33 lines (22 loc) · 1.65 KB

common_usecase.md

File metadata and controls

33 lines (22 loc) · 1.65 KB

Intro

This doc is about how these utils intended to be used.

Check hardware

Connect to the server that seems to be overwhelmed with network traffic. Run server-info --rate --subsystem to see if the server can process given traffic.

  1. If not - you'll see what is the biggest problem.
  2. You can see details on device level: server-info --rate --devices
  3. Or specify a subsystem: server-info --rate --net
  4. Or view details instead of rating server-info --show --memory
  5. If you see some device has low-performance - replacing it may be the quickest solution.

Check workload

If hardware is ok let's see what happening with current workload with command: network-top

  1. It highlights errors and problems with red or yellow color.
  2. It provide you information how load is distributed across the CPUs and network cards.
  3. It shows how devices attached to each other (highlights NUMA-bindings of CPUS and network cards). It can give an idea how to connect them for better cache-hits.

Configure network stack and hardware

If load is not distributed well or there still exist performance problems - you can tune it.

  1. Disable floating CPU-frequency, use maximize-cpu-freq command.
  2. Distribute interrupts of multi-queue network cards with rss-ladder <device> command.
  3. Apply autorps <device> for single-queue network cards.
  4. If everything is distributed well but there is still packet loss - increase buffers with rx-buffers-incrase <device>
  5. After each applied setting re-check state of network stack with network-top
  6. Add all commands you used to tune to an autostart script /etc/rc.local for example, otherwise it will work until reboot.