查看文章

Performance is a crucial issue in parallel/distributed applications. Designers and developers expect their applications to reach certain performance indexes to meet the expectations of high performance computing systems. Therefore, parallel application developers are obliged to not only to analyze the application itself, but also the software layers involved and the distributed system behavior, in order to detect any performance bottlenecks that appear during the execution of the application, determine their causes and modify the application to overcome such bottlenecks. So, performance analysis becomes a highly significant task in the development of parallel/distributed applications.

However, there is still a lack of geuinely useful tools and the most popular approach to carrying out performance analysis is the use of visualization tools [1–3] that offer to the developer a set of views of the application execution. The analysis of these views is a difficult and time consuming task that requires a high degree of expertise of the user. To tackle all these problems, user-friendly tools should be available. In this context several automatic performance analysis tools have been developed, Expert [4], Scalea [5]. These tools take a trace file from the execution of the application and try to detect performance bottlenecks using certain performance properties. The post-mortem approach has the advantage of being able to consider all detailed information gathered during application execution and, moreover, the analysis phase does not introduce any overhead in the application execution. Using this approach it is possible to perform a comprehensive analysis …