Friday, February 19, 2010

Optimizing the JBoss Microcontainer

Ales linked to the ongoing optimization work before I had planned on writing anything, so here is a very rushed overview of what is going on :-) 

For JBoss Application Server 6.0.0.M2 a lot of work went in to optimizing the boot time. Some of the work done by the AS team that really made a difference was optimizing the algorithm to add the deployers, and to make things like the admin console start when first accessed rather than use eager loading.

As part of this, for JBoss Kernel 2.2.0.Alpha6 I took a look at optimizing the dependency resolution algorithm of the Microcontainer, which is used to install and wire together the services the core of the application server. The dependency resolution algorithm is used to determine if the beans implementing the services can be moved through the states on the way from being put into the Microcontainer to being fully up and running. Initially I did work on a fully optimized prototype, but this was too big and risky an undertaking for the short timeframe and did not yield the expected results. This prototype is far from "finished" and there is plenty of room for improvement, so I might come back to it at some later stage.

So instead I went through and used JProfiler to see what is taking the time in the dependency resolution algorithm that we have today. What really made a difference here was that I found that when a bean can not be moved to the next state we were looking up its unresolved dependency three times in the Microcontainer. Wrapping these three calls into one gave a lot less overhead. That was the main optimization we had time for for JBoss Kernel 2.2.0.Alpha6 which is what is used in AS 6.0.0.M2. 

Since then I have been working on quite a few other optimizations that you can see on the Microcontainer Development Forum, more specifically here, which so far significantly improve the boot time of 6.0.0.M2 in my local benchmarks, and I am nowhere not done yet. The fixes are mainly making sure that objects used in maps are properly hashed, and making objects that are expensive to create less expensive to create. All these fixes will all go into the next community release of JBoss AS: 6.0.0.M3.

Another thing is that John Bailey is currently integrating JBoss Virtual File System 3 into AS. VFS is used as an abstraction layer to read files from several different sources, and VFS 3 promises to really speed things up when reading from zipped archives. This is something that happens a lot in the application server, when scanning archives for deployment descriptors and annotations, and loading classes and so on. So it will be very exciting to see this in JBoss AS 6.0.0.M3!