Re: [RFC PATCH 00/30] Code tagging framework and applications

[Kent Overstreet <kent.overstreet@xxxxxxxxx>]

On Fri, Sep 02, 2022 at 06:02:12AM -0600, Jens Axboe wrote:
> On 9/1/22 7:04 PM, Roman Gushchin wrote:
> > On Thu, Sep 01, 2022 at 08:17:47PM -0400, Kent Overstreet wrote:
> >> On Thu, Sep 01, 2022 at 03:53:57PM -0700, Roman Gushchin wrote:
> >>> I'd suggest to run something like iperf on a fast hardware. And maybe some
> >>> io_uring stuff too. These are two places which were historically most 
> >>> sensitive
> >>> to the (kernel) memory accounting speed.
> >>
> >> I'm getting wildly inconsistent results with iperf.
> >>
> >> io_uring-echo-server and rust_echo_bench gets me:
> >> Benchmarking: 127.0.0.1:12345
> >> 50 clients, running 512 bytes, 60 sec.
> >>
> >> Without alloc tagging:     120547 request/sec
> >> With:                      116748 request/sec
> >>
> >> https://github.com/frevib/io_uring-echo-server
> >> https://github.com/haraldh/rust_echo_bench
> >>
> >> How's that look to you? Close enough? :)
> > 
> > Yes, this looks good (a bit too good).
> > 
> > I'm not that familiar with io_uring, Jens and Pavel should have a better 
> > idea
> > what and how to run (I know they've workarounded the kernel memory 
> > accounting
> > because of the performance in the past, this is why I suspect it might be an
> > issue here as well).
> 
> io_uring isn't alloc+free intensive on a per request basis anymore, it
> would not be a good benchmark if the goal is to check for regressions in
> that area.

Good to know. The benchmark is still a TCP benchmark though, so still useful.

Matthew suggested
  while true; do echo 1 >/tmp/foo; rm /tmp/foo; done

I ran that on tmpfs, and the numbers with and without alloc tagging were
statistically equal - there was a fair amount of variation, it wasn't a super
controlled test, anywhere from 38-41 seconds with 100000 iterations (and alloc
tagging was some of the faster runs).

But with memcg off, it ran in 32-33 seconds. We're piggybacking on the same
mechanism memcg uses for stashing per-object pointers, so it looks like that's
the bigger cost.