Latest Jackrabbit performance test results

[Jukka Zitting <ju...@gmail.com>]

Hi,

Results from a performance test run from last night:

http://people.apache.org/~jukka/jackrabbit/report-2011-08-25/report.html

Some remarks:

* The BigFileReadTest results are pretty fascinating. It looks like
the set of ten 100MB test files fits into the OS buffer cache (the
test computer has 4GB RAM) and unlike in previous Jackrabbit versions
we can now for some reason avoid invalidating that cache. Or was there
some other recent DataStore change that I missed?

* We have a notable regression in concurrent read and read/write
performance. I'll file a blocker issue for that, to be fixed before
Jackrabbit 2.3.

* Also login performance is down, most likely due to extra access
control stuff. We had a similar issue before 2.2 when some cache setup
was taking a lot of time. Perhaps something similar is happening
again.

* Massive performance improvements for SetProperty and
UpdateManyChildNodes tests. I'm not sure which of the recent changes
is giving this performance boost.

* And as expected from JCR-2835, SQL2 descendant search performance is
way better than in 2.2 (note that we backported this change to 2.2.1).

BR,

Jukka Zitting

Re: Latest Jackrabbit performance test results

[Jukka Zitting <ju...@gmail.com>]

Hi,

On Thu, Aug 25, 2011 at 11:42 AM, Jukka Zitting <ju...@gmail.com> wrote:
> Results from a performance test run from last night:
>
> http://people.apache.org/~jukka/jackrabbit/report-2011-08-25/report.html

Here's an updated report:

http://people.apache.org/~jukka/jackrabbit/report-2011-09-01/report.html

> * The BigFileReadTest results are pretty fascinating. It looks like
> the set of ten 100MB test files fits into the OS buffer cache (the
> test computer has 4GB RAM) and unlike in previous Jackrabbit versions
> we can now for some reason avoid invalidating that cache.

That was indeed the case. I increased the test set from ten to a
hundred 100MB files, which prevented all of them from fitting into
memory at a time, and now there's much less variance in performance.

Note that there was a massive performance problem in writing large
files in the default configuration of all of Jackrabbit 1.4, 1.5 and
1.6. I had to disable the big file tests on those versions as
otherwise writing the 10GB of test data would have taken way too long
to be practical.

> * We have a notable regression in concurrent read and read/write
> performance. I'll file a blocker issue for that, to be fixed before
> Jackrabbit 2.3.

See https://issues.apache.org/jira/browse/JCR-3064

> * Massive performance improvements for SetProperty and
> UpdateManyChildNodes tests. I'm not sure which of the recent changes
> is giving this performance boost.

The performance boost can not be seen in the latest results anymore,
so it might have been just a fluke.

BR,

Jukka Zitting