Re: Created benchmarks modules for page_pool

[Jesper Dangaard Brouer <brouer@redhat.com>]

On Wed, 22 Jan 2020 12:42:05 +0200
Ilias Apalodimas <ilias.apalodimas@linaro.org> wrote:

> Hi Jesper, 
> 
> On Tue, Jan 21, 2020 at 05:09:45PM +0100, Jesper Dangaard Brouer wrote:
> > Hi Ilias and Lorenzo, (Cc others + netdev)
> > 
> > I've created two benchmarks modules for page_pool.
> > 
> > [1] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/lib/bench_page_pool_simple.c
> > [2] https://github.com/netoptimizer/prototype-kernel/blob/master/kernel/lib/bench_page_pool_cross_cpu.c
> > 
> > I think we/you could actually use this as part of your presentation[3]?  
> 
> I think we can mention this as part of the improvements we can offer,
> alongside with native SKB recycling.

Yes, but you should notice that the cross CPU return benchmark test
show that we/page_pool is too slow...


> > 
> > The first benchmark[1] illustrate/measure what happen when page_pool
> > alloc and free/return happens on the same CPU.  Here there are 3
> > modes of operations with different performance characteristic.
> > 
> > Fast_path NAPI recycle (XDP_DROP use-case)
> >  - cost per elem: 15 cycles(tsc) 4.437 ns
> > 
> > Recycle via ptr_ring
> >  - cost per elem: 48 cycles(tsc) 13.439 ns
> > 
> > Failed recycle, return to page-allocator
> >  - cost per elem: 256 cycles(tsc) 71.169 ns
> > 
> > 
> > The second benchmark[2] measures what happens cross-CPU.  It is
> > primarily the concurrent return-path that I want to capture. As this
> > is page_pool's weak spot, that we/I need to improve performance of.
> > Hint when SKBs use page_pool return this will happen more often.
> > It is a little more tricky to get proper measurement as we want to
> > observe the case, where return-path isn't stalling/waiting on pages
> > to return.
> > 
> > - 1 CPU returning  , cost per elem: 110 cycles(tsc)   30.709 ns
> > - 2 concurrent CPUs, cost per elem: 989 cycles(tsc)  274.861 ns
> > - 3 concurrent CPUs, cost per elem: 2089 cycles(tsc) 580.530 ns
> > - 4 concurrent CPUs, cost per elem: 2339 cycles(tsc) 649.984 ns  

Add a small bug, thus re-run of cross_cpu bench numbers:

- 2 concurrent CPUs, cost per elem:  462 cycles(tsc) 128.502 ns
- 3 concurrent CPUs, cost per elem: 1992 cycles(tsc) 553.507 ns
- 4 concurrent CPUs, cost per elem: 2323 cycles(tsc) 645.389 ns


> Interesting, i'll try having a look at the code and maybe run then on
> my armv8 board.

That will be great, but we/you have to fixup the Intel specific ASM
instructions in time_bench.c (which we already discussed on IRC).

> > 
> > [3] https://netdevconf.info/0x14/session.html?tutorial-add-XDP-support-to-a-NIC-driver


-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer