Hot IRQ CPU

[amnonh]

Scylla can use dedicafted cores for NIC IRQ processing, these cores can become a bottleneck.

Describe the condition when the issue happens

• CPU spent in softIRQ context is way higher in the dedicated CPU
• For how long:

What the user should do when this happens

• Add more “IRQ” CPUs
• Reduce the number of NIC’s HW queues - one CPU per ”IRQ” is more than enough:

Any Other info or Links.

[tzach]

@vladzcloudius how to extract this info?

[vladzcloudius]

@vladzcloudius how to extract this info?

The easiest way is to check the output of perftune.py and see which CPUs it binds NIC IRQs to.

There are other ways too but they would be depending on the current perftune.py implementation which may change in the future.