Re: Performance issues with vnet jails + epair + bridge

From: Sad Clouds <cryintothebluesky_at_gmail.com>
Date: Fri, 13 Sep 2024 07:36:59 UTC
On Fri, 13 Sep 2024 08:03:56 +0100
Sad Clouds <cryintothebluesky@gmail.com> wrote:

> I built new kernel with "options RSS" however TCP throughput performance
> now decreased from 128 MiB/sec to 106 MiB/sec.
> 
> Looks like the problem has shifted from epair to netisr
> 
>   PID USERNAME    PRI NICE   SIZE    RES STATE    C   TIME    WCPU COMMAND
>    12 root        -56    -     0B   272K CPU3     3   3:45 100.00% intr{swi1: netisr 0}
>    11 root        187 ki31     0B    64K RUN      0   9:00  62.41% idle{idle: cpu0}
>    11 root        187 ki31     0B    64K CPU2     2   9:36  61.23% idle{idle: cpu2}
>    11 root        187 ki31     0B    64K RUN      1   8:24  55.03% idle{idle: cpu1}
>     0 root        -64    -     0B   656K -        2   0:50  21.50% kernel{epair_task_2}

I think the issue may be to do with the genet driver itself. I think
the hardware is limited to one CPU per send or receive interrupt. On
Linux the best I can do is set SMP affinity for send on CPU0 and for
receive on CPU1, but that still leaves 2 other CPUs idle.

$ cat /proc/interrupts
           CPU0       CPU1       CPU2       CPU3       
...
 37:      74141          0          0          0     GICv2 189 Level     eth0
 38:      43174          0          0          0     GICv2 190 Level     eth0