Re: [PATCH v8 1/3] eal/x86: run-time dispatch over memcpy

[Thomas Monjalon <thomas@monjalon.net>]

25/10/2017 08:55, Li, Xiaoyun:
> From: Li, Xiaoyun
> > From: Richardson, Bruce
> > > On Thu, Oct 19, 2017 at 11:00:33AM +0200, Thomas Monjalon wrote:
> > > > 19/10/2017 10:50, Li, Xiaoyun:
> > > > > From: Thomas Monjalon
> > > > > > 19/10/2017 09:51, Li, Xiaoyun:
> > > > > > > From: Thomas Monjalon [mailto:thomas@monjalon.net]
> > > > > > > > 19/10/2017 04:45, Li, Xiaoyun:
> > > > > > > > > Hi
> > > > > > > > > > > >
> > > > > > > > > > > > The significant change of this patch is to call a
> > > > > > > > > > > > function pointer for packet size > 128
> > > (RTE_X86_MEMCPY_THRESH).
> > > > > > > > > > > The perf drop is due to function call replacing inline.
> > > > > > > > > > >
> > > > > > > > > > > > Please could you provide some benchmark numbers?
> > > > > > > > > > > I ran memcpy_perf_test which would show the time cost
> > > > > > > > > > > of memcpy. I ran it on broadwell with sse and avx2.
> > > > > > > > > > > But I just draw pictures and looked at the trend not
> > > > > > > > > > > computed the exact percentage. Sorry about that.
> > > > > > > > > > > The picture shows results of copy size of 2, 4, 6, 8,
> > > > > > > > > > > 9, 12, 16, 32, 64, 128, 192, 256, 320, 384, 448, 512,
> > > > > > > > > > > 768, 1024, 1518, 1522, 1536, 1600, 2048, 2560, 3072,
> > > > > > > > > > > 3584, 4096, 4608, 5120, 5632, 6144, 6656, 7168,
> > > > > > > > > > 7680, 8192.
> > > > > > > > > > > In my test, the size grows, the drop degrades. (Using
> > > > > > > > > > > copy time indicates the
> > > > > > > > > > > perf.) From the trend picture, when the size is
> > > > > > > > > > > smaller than
> > > > > > > > > > > 128 bytes, the perf drops a lot, almost 50%. And above
> > > > > > > > > > > 128 bytes, it approaches the original dpdk.
> > > > > > > > > > > I computed it right now, it shows that when greater
> > > > > > > > > > > than
> > > > > > > > > > > 128 bytes and smaller than 1024 bytes, the perf drops
> > > > > > > > > > > about
> > > 15%.
> > > > > > > > > > > When above
> > > > > > > > > > > 1024 bytes, the perf drops about 4%.
> > > > > > > > > > >
> > > > > > > > > > > > From a test done at Mellanox, there might be a
> > > > > > > > > > > > performance degradation of about 15% in testpmd
> > > > > > > > > > > > txonly
> > > with AVX2.
> > > > > > > > > >
> > > > > > > > >
> > > > > > > > > I did tests on X710, XXV710, X540 and MT27710 but didn't
> > > > > > > > > see
> > > > > > > > performance degradation.
> > > > > > > > >
> > > > > > > > > I used command "./x86_64-native-linuxapp-gcc/app/testpmd
> > > > > > > > > -c 0xf -n
> > > > > > > > > 4 -- -
> > > > > > > > I" and set fwd txonly.
> > > > > > > > > I tested it on v17.11-rc1, then revert my patch and tested it
> > again.
> > > > > > > > > Show port stats all and see the throughput pps. But the
> > > > > > > > > results are similar
> > > > > > > > and no drop.
> > > > > > > > >
> > > > > > > > > Did I miss something?
> > > > > > > >
> > > > > > > > I do not understand. Yesterday you confirmed a 15% drop with
> > > > > > > > buffers between
> > > > > > > > 128 and 1024 bytes.
> > > > > > > > But you do not see this drop in your txonly tests, right?
> > > > > > > >
> > > > > > > Yes. The drop is using test.
> > > > > > > Using command "make test -j" and then " ./build/app/test -c f -n 4 "
> > > > > > > Then run "memcpy_perf_autotest"
> > > > > > > The results are the cycles that memory copy costs.
> > > > > > > But I just use it to show the trend because I heard that it's
> > > > > > > not
> > > > > > recommended to use micro benchmarks like test_memcpy_perf for
> > > > > > memcpy performance report as they aren't likely able to reflect
> > > > > > performance of real world applications.
> > > > > >
> > > > > > Yes real applications can hide the memcpy cost.
> > > > > > Sometimes, the cost appear for real :)
> > > > > >
> > > > > > > Details can be seen at
> > > > > > > https://software.intel.com/en-us/articles/performance-optimiza
> > > > > > > ti
> > > > > > > on-of-
> > > > > > > memcpy-in-dpdk
> > > > > > >
> > > > > > > And I didn't see drop in testpmd txonly test. Maybe it's
> > > > > > > because not a lot
> > > > > > memcpy calls.
> > > > > >
> > > > > > It has been seen in a mlx4 use-case using more memcpy.
> > > > > > I think 15% in micro-benchmark is too much.
> > > > > > What can we do? Raise the threshold?
> > > > > >
> > > > > I think so. If there is big drop, can try raise the threshold. Maybe 1024?
> > > but not sure.
> > > > > But I didn't reproduce the 15% drop on mellanox and not sure how
> > > > > to
> > > verify it.
> > > >
> > > > I think we should focus on micro-benchmark and find a reasonnable
> > > > threshold for a reasonnable drop tradeoff.
> > > >
> > > Sadly, it may not be that simple. What shows best performance for
> > > micro- benchmarks may not show the same effect in a real application.
> > >
> > > /Bruce
> > 
> > Then how to measure the performance?
> > 
> > And I cannot reproduce 15% drop on mellanox.
> > Could the person who tested 15% drop help to do test again with 1024
> > threshold and see if there is any improvement?
> 
> As Bruce said, best performance on micro-benchmark may not show the same effect in real applications.

Yes real applications may hide the impact.
You keep saying that it is a reason to allow degrading memcpy raw perf.
But can you see better performance with buffers of 256 bytes with
any application thanks to your patch?
I am not sure whether there is a benefit keeping a code which imply
a signicative drop in micro-benchmarks.

> And I cannot reproduce the 15% drop.
> And I don't know if raising the threshold can improve the perf or not.
> Could the person who tested 15% drop help to do test again with 1024 threshold and see if there is any improvement?

We will test a increased threshold today.