Fork me on GitHub

n. Slang a rough lawless young Kuali developer.
[perhaps variant of Houlihan, Irish surname]
kualiganism n

Blog of an rSmart Java Developer. Full of code examples, solutions, best practices, et al.

Friday, April 23, 2010

Tuning the Garbage Collection in KFS

KFS does not ship with JVM configuration that optimizes the application. Implementing institutions are expected to do that themselves. At University of Arizona, we have spent some time tuning KFS garbage collection for what seems to be pretty decent performance. Our results came mostly from trial-and-error, research from literature, and Tuning Garbage Collection with the 5.0 Java[tm] Virtual Machine

This is what we've come up with.

-Xms2g -Xmx2g -XX:MaxPermSize=512m -Doracle.jdbc.Trace=true -Djava.util.logging.config.file=/usr/share/tomcat5/conf/ojdbclog.conf -server -XX:+UseParNewGC -XX:MaxNewSize=256m -XX:NewSize=256m -XX:SurvivorRatio=128 -XX:MaxTenuringThreshold=0 -XX:+UseTLAB -XX:+UseConcMarkSweepGC -XX:+CMSClassUnloadingEnabled -XX:+CMSPermGenSweepingEnabled

Concurrent Low Pause Collector

What we decided to use is the Concurrent Low Pause Collector. We run our application out of a 64-bit Redhat Enterprise Linux VM running from Vmware ESX Server with 2 processors and 4 gb of RAM. Now that's a mouthful. You may be wondering, "With 2 processors, why don't you use the parallel GC?" Well, we tried both parallel and concurrent low pause gc. Really, the only reason why you would use a parallel is not because you have an extra processor, but rather that sacrificing that processor during runtime is better than bringing the system to its knees. Well, there's got to be a way that you run the garbage collector without sacrificing processing power at what might be a crucial time. That's where the concurrent low pause collector comes in. Concurrent means it runs in a separate thread. You may sacrifice a processor here, yes. Low Pause means that unlike a Parallel Collector, this runs in a separate thread for a short amount of time. Win-win!! So what's this other stuff? Read on.

Heap Memory Requirements

We set the min and max to be the same. This way the JVM doesn't have to reallocate memory. It's greedy and takes everything it can right away. This saves time. In a server environment, you don't want to be conservative. Greedy is a better way to go.

Optimizing for Caching

The -XX:MaxNewSize=256m -XX:NewSize=256m properties grow the Eden gen space to 256 mb immediately. Just like with the heap, we don't want to reallocate on a server platform. Let's be as greedy as possible. We've optimized to 256mb because we expect to have a large amount of cache rewriting. Tuning JVM Garbage Collection for Production Deployments recommends setting it to 32mb for caching systems, but we've found that for larger retained caches like what we want, maybe 256mb is more about what we want to use. This is especially the case with the number and size of HashMap instances in use in KFS between Spring and Rice. We also optimize the survivor space for the young generation to be 128th the size of the eden space. MaxTenuringThreshold is turned off so that a new NewSize space becomes reusable with each collection. This is actually really really small, and works well for caching.

Optimizing for Concurrency

CMSClassUnloadingEnabled, CMSPermGenSweepingEnabled, and UseConcMarkSweepGC give us our concurrentcy collection. Together, they force different algorithms for the GC that are optimal for multiprocessor system like forcing class unloading to prevent the GC from being intrusive on the application. UseTLAB "uses thread-local object allocation blocks. This improves concurrency by reducing contention on the shared heap lock." from Tuning JVM Garbage Collection for Production Deployments.


We had tried numerous configurations, and this worked out the best for us. It gave us a 4x improvement when processing large batch jobs during the day. In most cases, we won't process large batch jobs while users are in the system; however, for testing we are limited on servers and sometimes we try to get more out of our systems than they can give.

No comments:

Post a Comment