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

[Thomas Monjalon <thomas@monjalon.net>]

02/11/2017 11:58, Li, Xiaoyun:
> From: Thomas Monjalon [mailto:thomas@monjalon.net]
> > 02/11/2017 11:22, Wang, Zhihong:
> > > > I don't know what is creating this drop exactly.
> > > > When doing different tests on different environments, we do not see
> > > > this drop.
> > > > If nobody else can see such issue, I guess we can ignore it.
> > >
> > > Hi Thomas, Xiaoyun,
> > >
> > > With this patch (commit 84cc318424d49372dd2a5fbf3cf84426bf95acce) I
> > > see more than 20% performance drop in vhost loopback test with testpmd
> > > macswap for 256 bytes packets, which means it impacts actual
> > > vSwitching performance.
> > >
> > > Suggest we fix it or revert it for this release.
> > 
> > I think we need more numbers to take a decision.
> > What is the benefit of this patch? In which use-cases?
> 
>  The benefit is that if compile it on a lower platform (such as only supports SSE),
> when it run on higher platforms (such as AVX2 or AVX512). It would still can get ISA benefit (AVX2).

Yes, but you don't provide any number here.

> User case seems to be that some customers want it in cloud environment and don't want to compile on all platforms.
> 
> > What are the drawbacks? In which use-cases?
> 
> The drawback is perf drop. So far, see lot of drop in mellanox case and vhost case.
> 
> Should I send the revert patch or you revert it directly?

You should send the revert yourself with some good justifications.
I did not ask some numbers when accepting the patch (my mistake).
Please provide the numbers for the revert.

> > Please, it is a call to test performance with and without this patch in more
> > environments (CPU, packet size, applications).

Who can test it in more environments?