PAMDA: Performance Assessment Using MAQAO Toolset and Differential Analysis

Identifying performance bottlenecks in applications is crucial to improve their efficiency, but it might be difficult to precisely assess their impact on performance: in particular, two performance problems can interact making it difficult to isolate and therefore to correct them. We propose PAMDA, a methodology to single out performance problems through hierarchical bottlenecks detection. Important potential performance issues are classified in a ‘Performance Breakdown Tree’ which is used to drive our iterative analysis cycle, prioritizing the most relevant problems. Our system relies on MAQAO toolset and code’s differential analysis. While MAQAO is a performance analysis and optimization tool suite, the differential analysis approach, which is implemented through DECAN tool, consists in quantifying performance changes when applying controlled transformations to the target code. Our focus will be on performance issues raised by processors and memory sub-systems in multicore architectures. We will demonstrate the approach on loops extracted from real life HPC applications.