[jboss-svn-commits] JBL Code SVN: r32146 - labs/jbosstm/workspace/whitingjr/trunk/performance.

[jboss-svn-commits at lists.jboss.org]

Author: whitingjr
Date: 2010-03-19 05:55:55 -0400 (Fri, 19 Mar 2010)
New Revision: 32146

Modified:
   labs/jbosstm/workspace/whitingjr/trunk/performance/README.txt
Log:
Added details about JCA pooling.

Modified: labs/jbosstm/workspace/whitingjr/trunk/performance/README.txt
===================================================================
--- labs/jbosstm/workspace/whitingjr/trunk/performance/README.txt	2010-03-19 09:55:17 UTC (rev 32145)
+++ labs/jbosstm/workspace/whitingjr/trunk/performance/README.txt	2010-03-19 09:55:55 UTC (rev 32146)
@@ -14,6 +14,9 @@
  The framework uses multi-threading to run tests concurrently and concurrency classes to
  co-ordinate the two iterations avoiding overlapping of asynchronous executions.
 
+ The testing environment uses the Microcontainer to provide the beans infrastructure to run the test.
+ Database resource connection pooling is provided by the JCA implementation by JBoss.
+
  The output from the profiling system is a comma separated file. This file is formatted with 
  the first line as the header. This format is easy to paste into a spreadsheet and manipulate
  into graphs. There is a template spreadsheet which is created to show the time taken and
@@ -62,4 +65,15 @@
       the connection is cached per thread.
 
 Known issues:
+  The original performance profiling project called Java Interactive Profiler (JIP) was integrated
+ with the profiling system. However the JIP profiler affected the performance results excessively
+ and was abandoned. Instead the 'wall to wall' time is captured for the 'steady state' run.
+ Integration with profilers has not been abandoned and others can be used. An alternative 
+ working profiler configuration is JProfiler. Ideally, at a later date another open source
+ profiler will be integrated.
 
+ The connection pooling performance is possibly a scalability bottleneck. Based on lock 
+ contention (100+ threads) around the pooling code. More investigation required to confirm
+ this. To overcome CPU saturation due to saturation have change the ManagedConnectionFactory
+ to bypass the pooling and instead cache the connection per thread.
+