Performance Analysis of High Volume/Low Latency Marketplace Systems

The aim of this this presentation is to discuss some challeng-ing issues in performance analysis and modeling of JAVA based high throughput and low latency computing systems for electronic trading. It is based on test results from sys-tems developed and operated by the Swedish company Cin-nober Financial Technologies. A very important perfor-mance measure for such systems is the response time or what is also called the door-to-door latency time, i.e. the time it takes from the moment a transaction enters the marketplace to the moment the corresponding response/confirmation is returned from the system. A main source of variations of the response time is the load of the system. The studied systems are JAVA/JVM based, which means that certain processes regularly have to do garbage collection. This will also contribute to the variations of the response time. The latency depends very much on the load characteristics, and the transaction flows are often bursty. Typically, the av-erage latency is given for a certain load, but the median is often a better measure since the underlying distribution, due to garbage collection and queuing effects, typically has a rather fat tail. Hence, it is important to quantify the cumulative distribution function in order to find more de-tailed performance measures such as quantiles. We have studied how to use a two mode Bernoulli queue model to model garbage collection service rate effects. It is possible to analyze the corresponding model under classical queueing theory assumptions, but due to the uncertain input distribu-tion and other real world constraints the results are for our case not reliable. Instead, simulation based queuing mod-els, derived and calibrated by test data, are used to predict performance under more realistic conditions. All JAVA/JVM vendors are currently developing and re-leasing new improved real-time garbage collection technolo-gies. Performance testing and analysis on a system level ∗This work was partially supported by the Swedish Researc