Presentation: Performance Beyond Throughput:An OpenJ9 Case Study

For the past 14 years I've been working on Java JIT compiler development at IBM. However, I am not your typical compiler guy because I don't focus that much on code generator or optimizer. Instead I deal with heuristics that determine what methods to compile, when to compile them and how much to optimize them. With Java, and any other languages that can benefit from a JIT compiler, the overhead of the compilation needs to be kept in check. You simply cannot afford to compile every single method at the highest optimization level; instead the JIT compiler needs to be quite selective. I've also worked on the design of various runtime aspects of the JVM like code cache management, metadata cache management or interpreter profiler to name a few. More recently I got interested in the behavior of the JVM in the cloud which is a challenging domain because the environment is more dynamic.

The motivation for the talk is twofold:

First is to increase awareness about performance metrics that are usually ignored, for instance footprint to give just an example. Many times I have read in online forums the opinion that footprint is not important because nowadays servers have plenty of memory. This may not be true in the cloud where big machines are carved into smaller VMs.

Second is to draw attention to OpenJ9, a new open source JVM than IBM has donated to the Eclipse Foundation. OpenJ9 works in conjunction with OpenJDK and we think it offers a few performance advantages over HotSpot JVM which is traditionally used by developers.

Due to the OpenJ9 connection, the talk should appeal primarily to developers and architects that use Java in their projects. However, the performance aspects I touch on should appeal to a broader audience because performance is a key ingredient of any software system.

This is tightly connected to the motivation for this talk. We want the audience to leave with two key aspects: First is that there are many facets to performance and, as a developer, you have to consider them all because you don't know what your customer values most, you cannot fully guess how people are going to use your product. Even if you do guess it right today, that may change tomorrow. Secondly, we want to spread the word about OpenJ9 and why, in the cloud, it could be a good alternative to HotSpot. We want people to try it, use for their projects, come up with suggestions regarding improvements and, why not, even contribute code to the OpenJ9 project. I'll also show how to hunt for performance issues using OpenJ9 performance tools.

For me it's about performance and more specifically it's about the balance between contradictory performance metrics. How do you make the right trade-offs without user guidance, without having to specify a long command line or go through a long and painful tuning process. How can the JVM adapt to the running environment especially when the environment is constantly changing from one minute to the next like in the cloud. The JVM (and other software application too) needs to become a lot smarter. It needs to collect information all the time about the environment, gather clues and change its behavior to match the environment.