JBoss benchmarking

[krish_borland]

I'm attempting to benchmark JBoss 4.0 against Borland Enterprise
Server (BES) 6.0.

I've been reading a couple of performance tuning articles/papers
on JBoss. However, I've some trouble figuring out if things apply
only to a particular 4.0 release, or a 3.2.x or a prior one.

Let me enumerate through what I have done in 4.0, and what I
am looking for/have missed (correct me if I am wrong):

(1) Benchmark uses only Commit Option B in entity beans.
This is default for JBoss 4.x (as well as BES). Will
try commit option A at a later stage. No special entry
needed for Commit Option B in Jboss. Clustering not
important for this benchmark.
(2) Jboss: Use of Instance-Per-Transaction Policy. I set
this in "jboss.xml"

<ejb-name>....</ejb-name>
<local-jndi-name>....</local-jndi-name>
<configuration-name>Instance Per Transaction CMP 2.x EntityBean</configuration-name>

(3) Jboss: No-select-before-insert (no class name needed?) in
"jbosscmp-jdbc.xml"

<ejb-name>....</ejb-name>
<entity-command name="no-select-before-insert" />

(4) PreparedStatementCache size
<local-tx-datasource>
...
<prepared-statement-cache-size>50</prepared-statement-cache-size>
</local-tx-datasource>
(5) Jboss: Read-ahead strategy on find: in "jbosscmp-jdbc.xml"

<ejb-name>...</ejb-name>
.....
<read-ahead>
on-find
</read-ahead>

(6) cache list max

<ejb-name>...</ejb-name>
.....
<list-cache-max>5000</list-cache-max>

(7) Disabled debug in: CachedConnectionManager
in "jbossjca-service.xml" (debug is false)

Most/many of the above are already the default in BES, and so
out-of-the-box performance & scalability is not really a major
pain. My initial tests on a big multi-CPU box (single JVM)
showed an out-of-the-box Jboss lagging behind by an order of
magnitude, but I have the performance difference down to a
50-100% difference now under the same JDK, identical tuning
params, etc.

Questions:
==========
(1) Are the above accurate?
(2) Any more perf influencing params I have missed?
(3) Any mechanism to limit the thread pool size
of the EJB Container?
(4) Anymore disabling of timers/stats gathering?
(5) SFSB passivation timeout (default?): disabling?
(6) .... ?

Thanks,

-krish

[bill.burke]

I'm not sure how well this is documented, but in JBoss 3.2.6 there is a left-join feature that allows you to define deep prefetching when loading a bean. Check out the DTD.

[krish_borland]

Ok, I will try this, but of course I am assuming that
your suggestions are valid for JBoss 4.0.x as well?

I'm not going to try to re-benchmark the numbers
against an older - uncertified version of JBoss (CTS).

If the left-join feature is all I have missed, let me know.
Alternately, take a look at Borland Enterprise Server's
CMP 2.0 "perf" example if you think anything could be
done to further improve performance.

-krish

[radl01]

Hi krishnan,

can you specify your test appliaction more ? I think that your test app is based on CMP because you are CMP Borland expert or i'm wrong ?

I did a few test with BES and JBoss and you are right taht the BES is faster but in my test it was only 10 or 15 %.

Jan