I've recently come across the Microsoft updated versions of the WSTest Web Services Benchmark and the .NET StockTrader Sample Application. They wasted no time bragging about the results :-)
Microsoft encourages people to download the benchmark kit and perform their own tests, so I did that. I will ignore the StockTrader App for now because is more complex to install and analyze, I will focus on the WSTest benchmark. The .NET/WCF results are very good and the guys at benchmark labs seem to really know their stuff. It's a pity that the benchmark choose to compare .NET/WCF against WebSphere, probably the most expensive, slow and cumbersome of all Java Application servers.
In the Java-Land there are faster solutions to choose from. I decided to implement my own version of the benchmark to verify just how fast or how slow can a Java implementation be.
The test is essentially a XML serialization/deserialization benchmark, so I picked the speedy JiBX as the framework for Java/XML data binding. JiBX is only as fast as the underlying XML parser and the fastest Stax parser that I know is the Aalto XML Processor. We will also need an HTTP layer and for this I really like the Mina Http Codec.
With all the ingredients in place it didn't took long to produce a benchmark implementation that doesn't suck :-). The code is available here.
And what about the results? Unfortunately I don't have a server class machine laying around for running proper tests. However, I do have VirtualBox and two virtual machines, one with "Windows2003 Server" that runs the "Self-Hosted" WSTest application and another with "Ubuntu 8.10 Server" that runs the Java implementation using the sun-java-jdk-1.6.10 jvm. Using soapUI as a load generator the "linux/java" setup runs circles around the "windows/.net/wcf", in some cases the throughput numbers are more than twice as high. Of course that these results should be taken with a truckload of salt. The tests should have been performed with a proper server machine, using Windows 2008 Server in the .NET setup and with several machines running the load generators. I would love to hear from someone that has a "benchmark lab".
Update: The http bits are now handled by Grizzly, the performance seems to be better.
Update: Check the follow-up post for a more detailed performance test.
A subscrição é anónima e gera, no máximo, um e-mail por dia.
. Alternative serialization...
. WSTest, some numbers - Up...
. A fast implementation of ...
. Receiving events/notifica...