db-derby-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Knut Anders Hatlen <Knut.Hat...@Sun.COM>
Subject Performance testing (was: Re: [VOTE] 10.5.1.1 release candidate)
Date Wed, 29 Apr 2009 10:17:43 GMT
Kathey Marsden <kmarsdenderby@sbcglobal.net> writes:

> I agree with Myrna that DERBY-4181 and DERBY-4075 are worrisome and
> would add to the list the increased likelihood of DERBY-4053 with 10.4
> client and 10.5 server and DERBY-4180 an apparent regression with 10.4
> client and 10.5 server.  It would have been nice to get the war file
> tested and do some performance testing too.

Regarding the performance testing, I haven't run any tests on the
release candidates, but Olav's nightly performance regression tests do
still run against trunk every night, and they haven't shown much that
should worry us for quite some time (except perhaps the usual high
variation between runs).

http://home.online.no/~olmsan/derby/perf/

I usually click the "Last year" links to get an impression of what's
variation and what's actually an improvement or a regression.

These tests give a pretty good picture of the performance of the general
execution machinery, but changes in BLOB/CLOB streaming performance
(where significant changes have been made in this release) and in the
optimizer will normally not be detected.

For this release, I've been able to notice two improvements by looking
at the tests:

* DERBY-3791 (Excessive memory usage when fetching small Clobs) had a
very positive effect on the Clob test (it did get backported to
10.4.2.0, so it will only be seen when upgrading from 10.4.1.3 or
earlier)

* DERBY-2991 (Index split deadlock) improved the performance when
running single-record select by primary key

On the negative side, I only found that single-record select using table
scan shows a slightly decreasing trend the last month or so, but only
for the configurations with a small number of concurrent threads:

http://home.online.no/~olmsan/derby/perf/select_sec_noindex_1y.html

I'll put a note on my todo list about checking whether this can be
attributed to any changes in the Derby code, or if it's just another
unpredicted effect of global warming or some other external factor.

-- 
Knut Anders

Mime
View raw message