Today, we are announcing the next generation server based on UltraSPARC T2 Plus processor and this server happens to be a monster in terms of performance. It can consolidate a lot of servers in a single box providing the similar combined throughput. For the server details and other blogs related to this server I suggest you to visit T5440 Blog index(http://blogs.sun.com/allanp/entry/sun_s_4_chip_cmt).
In this blog, I summarize the performance of IBM WebSphere Application Server (WAS) on T5440. To test and benchmark the performance, I use the most recent release WAS v7, that is based on Java EE 5 spec and JDK 6, with numerous features and performance enhancements. As I said earlier that this new Sun server is a monster in terms of performance, the WAS software set up and configuration require some planning and appropriate allocation of the system resources. When you do the psrinfo(1M) command on this server, it will report to you that the system has 256 "processors". This means that the processing power of this system far exceeds the software scalability of a single instance of the Application Server's. Thus, you will need multiple instances of WAS to drive the system to its full utilization. Solaris Containers provides the most efficient way to accomplish such configuration as Solaris Containers provides process space isolation among different WAS instance as well as allocating proper system resources.
For maximizing the utilization of the system, I configured the environment as follows.
- I created 7 Solaris Containers, allocated 32 processor threads to each of the 6 Containers, and allocated the remaining processor threads for the other container and the global zone. Then, I used the following WAS configuration:
- initialHeapSize="2500" maximumHeapSize="2500"
- -server -Xmn2048m -XX:+AggressiveOpts -XX:+UseParallelGC -XX:ParallelGCThreads=16 -XX:+UseParallelOldGC -Dcom.ibm.CORBA.TransportMode=Pluggable -Dcom.ibm.ws.pm.batch=true -Dcom.ibm.ws.pm.deferredcreate=true -Dcom.ibm.CORBA.FragmentSize=3000 -Dcom.ibm.ws.pm.useLegacyCache=false -Dcom.ibm.ws.pm.grouppartialupdate=true -Djavax.xml.transform.TransformerFactory=org.apache.xalan.processor.TransformerFactoryImpl -Djavax.xml.parsers.SAXParserFactory=org.apache.xerces.jaxp.SAXParserFactoryImpl -Djavax.xml.parsers.DocumentBuilderFactory=org.apache.xerces.jaxp.DocumentBuilderFactoryImpl -Dorg.apache.xerces.xni.parser.XMLParserConfiguration=org.apache.xerces.parsers.XML11Configuration
- Disable the PMI feature.
- Disable the System.out logging from the admin console if you want you can do the same by editing the config file. by editing the trace service like - startupTraceSpecification="\*=info:SystemOut=off"
- DynaCache - The DynaCache was used for this benchmark so the default size of the cahce was set to 2000 which wasn't enough for this test which can be adjusted accordingly based on the application need as I set it to 3000.
- For thread pool tuning the only thing that was relevant in this benchmark was tuning the WebContainer thread pool right so I did set it to 80 and which was more that enough and I think It can be scaled down a bit but havn't tried.
- Database connection pool was set it to same value as WebContainer pool and set the min and max to the same value.
For database I did not want to run into the network/disk contention so I created the DB in /tmp and I had to use two databases for this purpose. This was just to eliminated the Database configuration headaches and just measure the App Server machine scalability. These database were connected point-to-point with App Server box and the App Server instances were using the point-to-point connection for the DB.
This resulted in getting to 1.8X scalability of its predecessor system(T5140) in terms of throughput(which is measured for this benchmark in terms of requests served per second or more commonly known as req/sec).
So In nutshell if you are looking for another system to consolidate your WebSphere deployment with 1.8X throughput capacity of earlier UltraSparc T2 plus system then T5440 may be just the right system for you.