Updated and final results:
* svnadmin dump results now included, f6 / f7 is +/-20%
* fixed anomaly with ra_serf, results consistent with previous findings
'null-export' tests have been rerun and old results been replaced
* added a page on how 'null-export' reacts on cache configuration
to get a better picture of how cache size, ra layer and block-read
interact when caches are hot. Data for small or cold caches has
already been in covered in other tests.
-- Stefan^2.
On Mon, Jul 14, 2014 at 1:54 AM, Stefan Fuhrmann
<stefan.fuhrmann_at_wandisco.com> wrote:
> After almost a week of continuous measurements and 4320
> individual data point, the results are in now. Summary:
>
> * on cold disks, 'svn log -v' is 2x..3x as fast with packed f7 than with
> packed f6 in default config and up to 8x with advanced options
> * on cold disks, 'svn export' with packed f7 is 40%.50% faster in
> default config and 2x as fast with advanced options
> * ra_serf has an anomaly making "fast" config up to 100x slower
> than ra_svn - independent of repo format
>
> Details to how I set up the repositories and measurements plus
> the rationale behind it can be found in our wiki:
>
> https://wiki.apache.org/subversion/MeasuringRepositoryPerformance
>
> Tools and test script have been committed in r1610264. System
> setup is close to what Ivan used in his tests, except for having
> 4GB to accommodate the much larger BSD repo as well.
>
> Original measurement log and detailed results can be found in the
> .zip file attached. The spreadsheet has been upgraded and now
> displays f6/f7 because "+400%" is easier to understand than "-80%".
> It now also highlights data cells with particularly high variation.
>
> I'm currently waiting for the 'svnadmin dump' results to come in.
> Those will take about 3 days but the expectation is that packed f7
> may be slightly slower here due to the "unusual" access pattern.
> After that, I'll try to isolate the trigger for the ra_serf anomaly.
>
> -- Stefan^2.
Received on 2014-07-23 15:20:22 CEST