[Gluster-devel] Regular Performance Regression Testing

[Vijay Bellur]

On Mon, Aug 29, 2016 at 8:42 AM, Niels de Vos <ndevos at redhat.com> wrote:
> On Mon, Aug 29, 2016 at 05:55:00PM +0530, Nigel Babu wrote:
>> On Mon, Aug 29, 2016 at 01:49:52PM +0200, Niels de Vos wrote:
>> > On Mon, Aug 29, 2016 at 05:01:18PM +0530, Nigel Babu wrote:
>> > > Hello folks,
>> > >
>> > > I've had chats with Manoj and Ambarish about performance testing and what we
>> > > can do upstream. Niels today solved half my problem by pointing out that we can
>> > > get physical nodes on CentOS CI. The general idea is to run iozone[1] and
>> > > smallfile[2] on a fixed frequency for master (to begin with).
>> > >
>> > > Does this sound like a good idea? If so, read on.
>> > >
>> > > For this to happens a few things need to happen:
>> > > * I'll need some help from a few people who can read the reports and coordinate
>> > >   fixes. That is, someone needs to "own" performance for upstream.
>> > > * I need some help in generating the right reports so we can figure out if our
>> > >   performance went up or down.
>> >
>> > The provisioning in the CentOS CI does not allow us to select certain
>> > systems (yet). So you would get different performance results, depending
>> > on the hardware that the reservation request returns:
>> >   https://wiki.centos.org/QaWiki/PubHardware
>> >
>> > Also, these physical machines do not have additional disks. The single
>> > SSD that these systems have, is completely used by the installation, no
>> > free space to partition to our liking, no additional disks available.
>> >
>> > I welcome any additional testing that we can run regulary, but to call
>> > it 'performance testing' might be a little pre-mature. At least the
>> > performance results should be marked as 'unoptimized' or similar.
>> >
>> > HTH,
>> > Niels
>> >
>>
>> The goal of this testing, to begin with, wouldn't be to get absolute numbers
>> but to try and catch decrease in performance, if that makes sense. In essence,
>> it's regression testing but for performance.
>
> Ah, ok, changing the subject to reflect that.
>
>> Thank you for raising the fact that it may be inconsistent, I'll talk to the
>> Centos CI folks and see what's the best way forward for us before we get here.
>> But let's work with the assumption that I'll sort out the infra side of things.
>
> There has been a request from others already to be able to select the
> blade-chassis during reserving machines. This would come a long way to
> get comparable results. Otherwise we can compare the results based on
> sub-domain (per chassis) where the tests were running (more difficult
> when multiple machines/chassis are involved).
>


+1 to running performance regression tests on identical hardware.

I would also recommend running perf-test.sh [1] for regression.

Thanks,
Vijay

[1] https://github.com/avati/perf-test/blob/master/perf-test.sh