Re: [ovirt-devel] [VDSM] All tests using directio fail on CI
On Thu, Sep 29, 2016 at 12:55 PM, Anton Marchukov <amarchuk(a)redhat.com>
wrote:
> > The node will be offline for now until we agree on what to do.
> > An option is to abandon RAM disks completely as we didn't find
> > any performance benefits from using them so far.
>
> That's very surprising. On my case it doubles the performance, at least.
> But I assume my storage (single disk) is far slower than yours.
>
What amount of RAM you had available to Linux file system cache and were
there any previous runs so Linux were able to put any mock caches into the
RAM cache?
I don't do mock. And if I run everything in RAM (whether directly under
/dev/shm/<somewhere> or in a zram disk), I honestly don't need the Linux
system cache.
Besides the possible difference in disk speeds I think the second factor
is this Linux fs cache that basically create an analog of RAM disk on the
fly.
Well, theoretically, if you have enough RAM and you keep re-running, many
of the data is indeed going to be cached. I'd argue that it's a better use
of RAM to just run it there.
Those two things might explain why we do not see any performance
improvement from RAM drives in our case.
Indeed.
Y.
Anton.
Attachments:
- attachment.html (text/html — 2.6 KB)