launchpad-dev team mailing list archive

[Martin Pool <mbp@xxxxxxxxxxxxx>]

On 8 February 2011 21:54, Robert Collins <robertc@xxxxxxxxxxxxxxxxx> wrote:
> On Tue, Feb 8, 2011 at 10:22 PM, Julian Edwards
> <julian.edwards@xxxxxxxxxxxxx> wrote:
>> On Tuesday 08 February 2011 02:39:02 Robert Collins wrote:
>>> Mid end desktops these days are 8 way machines that should be able to
>>> easily do a full test run in < 60 minutes.
>>
>> Good luck with that.  I've parallelised test runs using virtualisation before
>> and you get so much I/O contention that the overall increase in speed is only
>> about 10%.
>>
>> (Using two simultaneous runs on my quad-core Phenom)
>
> I'm certainly aware that there are a bunch of of issues - and previous
> experiments here. One of the fairly early steps has to be generating a
> load profile for each test process - (python, librarian, appserver,
> postgresql etc). We need to then figure out answers to get enough
> capacity to scale. For instance, IO load can perhaps be largely
> addressed via running the test DB's in a ramdisk. I remember prior
> attempts with ramdisks having little impact - but that may can easily
> be accounted for if we consider that in a single threaded
> configuration we are not bottlenecked on IO: eliminating a
> non-bottleneck won't have a dramatic effect. In fact if we're
> sufficiently far below the IO thresholds in a single threaded test
> config, its plausible we wouldn't see any improvement.
>
> As a datapoint, bzr, which has parallelised its tests for some years
> now, scales nearly linearly - and its very IO heavy, writing GB's of
> data across an entire test run. If you hit the IO ceiling on a
> machine, it stops improving rapidly ;)

Yes, big thing to do with bzr is to put TMPDIR onto a tmpfs, which
will both give you a constant-factor improvement, and also much higher
scaling headroom.  bzr writes a lot of data, but also deletes it soon
afterwards so it's ideal for a ramdisk.  Launchpad probably won't get
quite that behaviour, but it might get close.

Martin