Abstract

When performance tools are used to analyze an application with thousands of processes, the data generated can be bigger than the memory size of the cluster node, causing this data to be loaded in swap memory. In HPC systems, moving data to swap is not always an option. This problem causes scalability limitations that affect the user experience and it presents serious restrictions for executing on a large scale. In order to obtain knowledge about the application’s performance, the performance tools usually instrument the application to generate the data. When the instrumented parallel application is executed with thousands of processes, the data generated may be higher than the memory size of the compute node used to analyze the data in order to obtain the knowledge. Performance tools such as PAS2P predict the execution time in target machines. In order to predict the performance, PAS2P carries out a data analysis with the data in each application process. The data collected is analyzed sequentially, which results in an inefficient use of system resources. To solve this, we propose designing a parallel method to solve the problem when we manage a high volume of data, decreasing its execution time and increasing scalability, improving the PAS2P toolkit to generate performance knowledge defined by the application’s behavior phases.