Re: [RFC PATCH 0/5] Fun with the multipathing code

[Olga Kornievskaia <aglo@umich.edu>]

On Wed, Jan 9, 2019 at 3:38 PM Trond Myklebust <trondmy@hammerspace.com> wrote:
>
> Hi Olga
>
> On Wed, 2019-01-09 at 14:39 -0500, Olga Kornievskaia wrote:
> > Hi Trond,
> >
> > Do you have any plans for this patch set?
> >
> > I applied the patches on top of 4.20-rc7 kernel I had and tested it
> > (linux to linux) with iozone on the hardware (40G link with Mellanox
> > CX-5 card).
> >
> > Results seem to show read IO improvement from 1.9GB to 3.9GB. Write
> > IO
> > speed seems to be the same (disk bound I'm guessing). I also tried
> > mounting tmpfs. Same thing.
> >
> > Seems like a useful feature to include?
>
> Thanks for testing this.
>
> Was this your own port of the original patches, or have you taken my
> branch from
> http://git.linux-nfs.org/?p=trondmy/linux-nfs.git;a=shortlog;h=refs/heads/multipath_tcp
> ?

I didn't know one existed. I just took original patches from the
mailing list and applied to 4.20-rc7 (they applied without issues that
I recall).

> Either way I appreciate the data point. I haven't seen too many other
> reports of performance improvements, and that's the main reason why
> this patchset has languished.
>
> 3.9GB/s would be about 31Gbps, so that is not quite wire speed, but
> certainly a big improvement on 1.9GB/s.

Maybe it's the lab setup that's not tuned to achieve max performance.

> I'm a little surprised, tbough,
> that the write performance did not improve with the tmpfs. Was all this
> using aio+dio on the client?

It is what ever "iozone -i0 -i1 -s52m -y2k -az -I" translates to.

To clarify by "didn't improve" I didn't mean the write speed with disk
is same as write speed with tmpfs (disk write speed is ~168MB and
tmpfs write speed is 1.47GB). I meant that it seems with nconnect=1 it
achieves the "max" performance of disk/tmpfs.

>
> --
> Trond Myklebust
> Linux NFS client maintainer, Hammerspace
> trond.myklebust@hammerspace.com
>
>