Java VMs Compared
Our current stack is based on phoneME Advanced, a configuration not available in Open Embedded. That forced us to look at other options as we started working with Poky. The original prototype of the BUG used JamVM and and GNU Classpath, but we have been using phoneME for quite a while. Now was a perfect opportunity to look at other VMs. The obvious alternative is JamVM, but we also tested the Cacao virtual machine due to its support by the Jalimo and OpenMoko projects.
Ken announced earlier this week that the next software release for the BUG includes some big changes. Many will be immediately visible to the user, but some of the biggest changes are coming under the hood. The new system is significantly more flexible allowing us to change the packages in our Linux distribution easily.
It seems not many people have been comparing VMs on ARM recently so some benchmarks were in order. We based our tests on this article on benchmarking. Our three benchmarks were Scimark2, Linpack, and startup time comparison.
First up was Scimark2, the most comprehensive of the benchmarks. The suite runs various algorithms and reports these scores in mega-flops per second. It also makes a composite score of these tests. JamVM is the clear winner. While Cacao has solid performance on the Monte Carlo prime number finding algorithm, it falls behind in all of the rest of the tests. PhoneME seems to be about 10-20% slower than JamVM in most tests.
Next up was the Linpack benchmark. The first graph is a measure of the speed of the processor in mega flops per second. The second graph is the overall time for running the benchmark so lower is better. Again here we see the same order as the Scimark2 tests. JamVM runs on top for both tests followed by phoneME and Cacao.
Finally, our startup time test. Starting up the OSGi stack currently takes a few seconds, so startup time of the VM has a significant impact on the overall boot time. For this benchmark we started up the Concierge OSGi framework - though it did not load any bundles since we had not ported all of them to the new build system at the time.
Here JamVM performed as expected with an impressive startup time. Note that this is not a constant time, these gains will increase with a larger load. When starting up our full OSGi stack you will probably gain a few seconds with JamVM.
All of this makes it look like the best choice is JamVM, but there are a lot of variables that are swept under the rug above. Most importantly, a 20% difference on a synthetic benchmark does not indicate a 20% improvement in end user experience.
Second, there are some differences in the configurations. We ran these tests using GNU Classpath on JamVM and Cacao. Keep in mind that GNU Classpath is J2SE-like in size, while phoneME Advanced is significantly lighter. This will make a big difference on a benchmark that measures full system performance. These benchmarks are designed for CPU throughput so the additional overhead of Classpath does not show up. However, both Cacao and JamVM can run with phoneME rather than Classpath, it is just not the default behavior in OpenEmbedded.
Perhaps the biggest red herring stems from us not building phoneME with JIT right now. This is due to a problem with older versions of the Linux kernel. With the new build system we will build phoneME with JIT which will gain significant performance. These benchmarks are comparing the current build of phoneME to the Cacao and JamVM default setups in OpenEmbedded, we are not comparing the maximum performance that can be squeezed out of each VM.
Finally, on Thursday Matthias Klose was able to get OpenJDK built for ARM. We will have to take a look at its performance as soon as we have it running on the BUG.
We still are very excited to see the community based VMs performing so well. While there are more factors to consider when choosing a JVM implementation (such as TCK validation, branding concerns, and community support) JamVM definitely warrants a look for anyone interested in embedded Java development on ARM based systems, it certainly impressed us. Look here for a follow up post sometime in the coming months.