Re: [GIT PULL v2] Networking for 6.7

[Alexei Starovoitov]

On Thu, Nov 9, 2023 at 10:09 AM Yonghong Song <yonghong.song@xxxxxxxxx> wrote:
>
>
> On 11/9/23 8:14 AM, Kirill A. Shutemov wrote:
> > On Thu, Nov 09, 2023 at 08:01:39AM -0800, Alexei Starovoitov wrote:
> >> On Thu, Nov 9, 2023 at 7:49 AM Kirill A. Shutemov <kirill@xxxxxxxxxxxxx> wrote:
> >>> On Tue, Oct 31, 2023 at 02:09:48PM -0700, Jakub Kicinski wrote:
> >>>>        bpf: Add support for non-fix-size percpu mem allocation
> >>> Recent changes in BPF increased per-CPU memory consumption a lot.
> >>>
> >>> On virtual machine with 288 CPUs, per-CPU consumtion increased from 111 MB
> >>> to 969 MB, or 8.7x.
> >>>
> >>> I've bisected it to the commit 41a5db8d8161 ("bpf: Add support for
> >>> non-fix-size percpu mem allocation"), which part of the pull request.
> >> Hmm. This is unexpected. Thank you for reporting.
> >>
> >> How did you measure this 111 MB vs 969 MB ?
> >> Pls share the steps to reproduce.
> > Boot VMM with 288 (qemu-system-x86_64 -smp 288) and check Percpu: field of
> > /proc/meminfo.
>
> I did some experiments with my VM. My VM currently supports up to 255 cpus,
> so I tried 4/32/252 number of cpus. For a particular number of cpus, two
> experiments are done:
>    (1). bpf-percpu-mem-prefill
>    (2). no-bpf-percpu-mem-prefill
>
> For 4 cpu:
>     bpf-percpu-mem-prefill:
>       Percpu:             2000 kB
>     no-bpf-percpu-mem-prefill:
>       Percpu:             1808 kB
>
>     bpf-percpu-mem-prefill percpu cost: (2000 - 1808)/4 KB = 48KB
>
> For 32 cpus:
>     bpf-percpu-mem-prefill:
>       Percpu:            25344 kB
>     no-bpf-percpu-mem-prefill:
>       Percpu:            14464 kB
>
>     bpf-percpu-mem-prefill percpu cost: (25344 - 14464)/4 KB = 340KB
>
> For 252 cpus:
>     bpf-percpu-mem-prefill:
>       Percpu:           230912 kB
>     no-bpf-percpu-mem-prefill:
>       Percpu:            57856 kB
>
>     bpf-percpu-mem-prefill percpu cost: (230912 - 57856)/4 KB = 686KB
>
> I am not able to reproduce the dramatic number from 111 MB to 969 MB.
> My number with 252 cpus is from ~58MB to ~231MB.

Even 231MB is way too much. We shouldn't be allocating that much.
Let's switch to on-demand allocation. Only when bpf progs that
user per-cpu are loaded.