Terracotta Discussion Forums (LEGACY READ-ONLY ARCHIVE)

how can i reduce terracotta overhead so i get the best performance?
say i'm using sun java 1.5, don't use annotations, don't use reflection or AOP etc.
is there a how-to on this issue?

http://www.terracotta.org/confluence/display/docs1/Tuning+Guide

Requires you to login....
thanks

You'll want to reduce the instrumentation. Most of the sample Terracotta configurations ship with a wildcard "instrument the world" instruction : to instrument *..*.

This works well to get your application running, but it is typically unnecessary to instrument every class.

Read this thread for some more details on what you should do in this case:

http://forums.terracotta.org/forums/posts/list/387.page

We'll update the Tuning Guide to include this kind of information.

As far as POJO clustering is concerned (i.e. not HttpSession clustering), you can also look at the lock granularity (the number of changes you make in a synchronized block) and partitioning of large data sets.

Each synchronized block of code (on shared objects) is one Terracotta Transaction. In some cases, if you can batch more application transactions per Terracotta transaction, you will gain in performance.

If you have a large data set, you can break down a map into map of maps.

In any case, for HttpSessions and POJOs, you should try and keep as much locality of reference as you can (e.g. through sticky load balancing). The less you cross reference shared objects (referencing the same object on more than one nodes), the better.