Spring/Hibernate Initialization Slowness

09 Jan 2006

As I commented in an earlier post, we’ve been transitioning our legacy EJB-based system to a new platform built largely on Spring and leverging hibernate for persistence.

The technology has worked well, I’ve seen improvements by way of code reduction and ease of implementation in certain areas. Instead of 500 remote methods exposed via 30-odd session beans (and stateful ones at that for a reason I won’t get into now), I’ve implemented a fairly straight forward command pattern variation that combines a metadata layer with EJB/MDB and Spring execution end-points. Persistence is easier, and a custom DSL built-on top of HQL has allowed us to replace nearly all of our EJBQL and CMP-code generically.

However, all is not bliss. We’re suffering from a bit of technical debt. We attempted somewhat of a model-driven approach with effort put into developing a UML model of our existing schema (and some sequence/class diagrams). The original goal was to use andromda tags to annotate our model and produce useful hibernate mapping files. It sounded good in practice, but of course it didn’t work as expected (and really wasn’t that big of a surprise either). We’ve annotated a lot of the model, but there are a number of hibernate features not supported by andromda that required some post-processing tweaking in Perl.

The other significant drawback I’ve noticed is slowness at initialization time. I know that we’ve introduced a lot of overhead in the parsing and interpretation of our own metadata layer, but there appears to be sufficient overhead in the initialization and preparation of the 130-odd Hibernate mapping files and 40-50 Spring beans. Disabling cglib has not shown much in the way of improvement, but we’re still investigating and profiling.

The last problem is one that was faced even in the old legacy system. Every 5 or 6 deploys we get an out of memory exception in JBoss. In the legacy system it was a regular OutOfMemoryException, in the new hibernate/spring platform (still deployed to JBoss with an EJB3 SLSB and MDB) we get out of memory problems in the Permgen space. I’m not the only one with this problem and it’s affecting all of our developers. I’ve bumped up the memory allocated to permgen to 128MB but that only delays the problem.

If anyone has any suggestions or ideas, feel free to comment.