New Analysis Techniques in the CEPBA-Tools Environment

Abstract

The CEPBA tools environment is a performance analysis environment that initially focused on trace visualization and analysis. Current development efforts try to go beyond the presentation of simple statistics by introducing more intelligence in the analysis of the raw data.

The paper presents an overview of three recent developments in this area. First, we show how spectral analysis techniques can be used to isolate sufficiently small regions of a trace that characterize the behavior of the whole run. Second, we describe how clustering analysis techniques can be used to identify temporal and spatial structure in parallel programs, an essential component to ease the job of the analyst, but also to automatically derive a broad range of both precise and focused metrics from a single run of a program. Then we describe how sampling and tracing data acquisition techniques can interoperate to generate with very low overhead extremely precise metrics about the temporal behavior of a program.

The development rests upon the trace based CEPBA-Tools environment, using the Paraver visualization capabilities to check the quality and usefulness of the techniques. Once identified, they can be implemented on-line aiming at maximizing the amount of information obtained from a run. We report the work being done on top of MRNET in this direction.

We consider that by applying and combining these and other techniques from various data analysis and mining fields, performance analysis tools will be able to effectively address the huge challenge posed by future exascale systems.