LPC Traffic Shaping w/ BPF Talk - percpu followup

LPC Traffic Shaping w/ BPF Talk - percpu followup

[Dennis Zhou]

Hi Eddie, Vlad, and Willem,

A few people mentioned to me that you guys were experiencing issues with
the percpu memory allocator. I saw the talk slides mention the
following two bullets:

1) allocation pattern makes the per cpu allocator reach a highly
   fragmented state
2) sometimes takes a long time (up to 12s) to create the PERCPU_HASH
   maps at startup

Could you guys elaborate a little more about the above? Some things
that would help: kernel version, cpu info, and a reproducer if possible?

Also, I did some work last summer to make percpu allocation more
efficient, which went into the 4.14 kernel. Just to be sure, is that a
part of the kernel you guys are running?

Thanks,
Dennis

Re: LPC Traffic Shaping w/ BPF Talk - percpu followup

[Vlad Dumitrescu]

On Tue, Nov 20, 2018 at 2:17 PM Dennis Zhou <dennis@kernel.org> wrote:
>
> Hi Eddie, Vlad, and Willem,
>
> A few people mentioned to me that you guys were experiencing issues with
> the percpu memory allocator. I saw the talk slides mention the
> following two bullets:
>
> 1) allocation pattern makes the per cpu allocator reach a highly
>    fragmented state
> 2) sometimes takes a long time (up to 12s) to create the PERCPU_HASH
>    maps at startup
>
> Could you guys elaborate a little more about the above? Some things
> that would help: kernel version, cpu info, and a reproducer if possible?
>
> Also, I did some work last summer to make percpu allocation more
> efficient, which went into the 4.14 kernel. Just to be sure, is that a
> part of the kernel you guys are running?
>
> Thanks,
> Dennis

Hi, Dennis,

Thanks a lot for reaching out and sorry for the delay in answering. I
was trying to build something which shows the problem on a recent
upstream kernel, but I was unable to do so until now.

It seems like I can still reliably reproduce on one of our kernels,
which has a lot of 'percpu: *' patches by you from 4.14, and with an
internal application. Unfortunately, we haven't spent too much time on
this issue, and I will have 'page in' state from a few months back.
Hopefully, I'll be able to reproduce on a vanilla kernel and a simpler
application, that we can publish, in the following days (after TG
weekend).

Thanks,
Vlad