Techniques for maximizing information and minimizing trace size

Speaker: Germán Llort (BSC)
Date: Friday, 4 December 2015, 15:30-16:45
Session: Performance Tools II
Talk type: Project talk (30 min)

Abstract: Performance analysis tools are essential not only to understand the behavior of parallel applications, but to identify why performance expectations might not have been met, serving as guidelines to improve the inefficiencies that caused poor performance, and driving both software and hardware optimizations. However, detailed analysis of the behavior of a parallel application requires to process a large amount of data that also grows extremely fast. Current large scale systems already comprise hundreds of thousands of cores, and upcoming Exascale systems are expected to assemble more than a million processing elements. With such number of hardware components, the traditional analysis methodologies consisting in blindly collecting as much data as possible and then performing exhaustive lookups are no longer applicable, because the volume of performance data generated becomes absolutely unmanageable to store, process and analyze. The evolution of the tools suggests that more complex approaches are needed, incorporating intelligence to perform competently the task of detailed analysis. To this end, our work leverages techniques that have been successfully applied to aid in the task of Big Data Analytics in fields like machine learning, data mining, signal processing and computer vision. We rely on these techniques to improve the analysis of large-scale parallel applications by automatically uncovering repetitive patterns, finding data correlations, detecting performance trends and further useful analysis information. Combining their use, we have been able to minimize the volume of performance data captured from an execution, while maximizing the information and insight gained from this data.