Enhanced Method for the Signature Adaptability to the Application Behavior

Predicting performance of parallel applications using benchmarks is becoming more complex. Each parallel application has different behavior and it is difficult to know which is the benchmark that most closely resembles the application, therefore the benchmarks have a low representativeness of real programs. We are proposing the users to predict the performance of their applications by generating a signature using a method called Parallel Application Signature for Performance Prediction (PAS2P), which represents the core of their applications. We are presenting an enhanced method for the signature adaptability to the application behavior. Until now, phases were considered independently, each one had a checkpoint, but now we analyze whether they are contiguous and can form a group. This is done by decreasing the storage needed for the signature, slightly increasing the signature execution time in order to predict the performance of the applications thus compensating the execution time of the signature and its size (MB). We show results from building different signatures for the same applications to allow the user to select between different options in order to create the signature.