Friday, September 18, 2009

Web Services Performance Tidbits

My team was recently working with a customer on their web services application and we wanted to document a few things we learned during the engagement.

Network bottlenecks

In our IBM labs, when we performance test web services, we set up a set of multi-threaded clients pounding a separate server system until the CPU utilization on the server system becomes saturated (that way - we can see where the time is spent processing). Working with the customer, they could only stress their system to a small portion of the CPU processing limit.

In the end, we discovered that the network was only a 100MB network and totally inadequate to handle the type of loads used in the test - whereas, in our IBM labs, we run on 1GB Ethernet. Using the size of the messages, we computed what the maximum achievable throughput would be on a 100MB network and these matched almost exactly to what was being reported for the customer. This showed us that the network was the bottleneck in the environment and the Application Servers are not being stressed at all - only the network connection.

One way to get around the network bottleneck in an interoperable way would be to use GZIP compression to compress the data (since the network is the bottleneck). See here and here for more information about enabling GZIP compression for Web Services.

Transaction Isolation Level

Another interesting point that we learned was that different Application Server vendors use different transaction isolation levels on the JDBC connection to the database. WebSphere, by default, uses Repeatable Read as a default isolation level which initially made our performance slower. This default was made to protect customers data. Another vendor used a default of Read Committed isolation level which offers higher throughput but more risk. The choice of which isolation level to use is up to the customer but in a comparison, this setting should be consistent to get apples-to-apples measurements.


WebSphere does provide many parameters that can be tuned to increase performance. Some of the important ones for web services are JVM heap size and Web container thread pools. A white paper was recently published by our performance team that shows the impact of adjusting some of these. See WAS Perf Tuning for more info

We hope this information helps others as they try to build performant web service applications.

1 comment:

  1. I've updated the WebSphere Performance Tuning link. Since the original post, an article came out on developerWorks - and the Performance Tuning document was moved to point to that article.