Serviceability under heavy loads

[krizhanovsky]

The other possible issue is that due to HTTP processing in softirq one can experience SSH unavailability under heavy traffic loads, even on designated network interface. It seems we should prioritize ingress traffic depending on service (SSH, HTTP etc.) using XDP or Linux QoS.

Linked with #488 if Linux QoS is used to fix the problem or #1048 if XDP is used.

Probably some scheduling ideas from Shenango can be applied to solve the issue.

[krizhanovsky]

Actually, DPDK has quite the same problem with separation data and control CPUs. In case of separate NICs it's possible to assign interrupt queues to different CPUs, completely or in some ratio. E.g. if there are 2 NICs with 64 queues each and 8 CPUs, then the data NIC can assign 9 queues to 7 data CPUs, 1 queue to the 8th, while the control NIC assigns all the queues to the 8th CPU.

There is no need to develop specific code, rather just write a Wiki best practice.