The MPP Apprentice™ Performance Tool: Delivering the Performance of the Cray T3D®

Abstract

The MPP Apprentice™ performance tool is designed to help users tune the performance of their Cray T3D, applications. By presenting performance information from the perspective of the user’s original source code, the MPP Apprentice tool helps users rapidly identify the location and cause of the most significant performance problems. Unlike many trace-based performance tools, the data collection mechanism permits fine-grained performance statistics to be collected with a low level of intrusion for work sharing, data parallel, and message passing codes. The low level of intrusion also permits the mechanism to scale to permit performance analysis on long running codes on thousands of processors. Information displayed within the tool includes total time through regions of code, instruction counts, time spent in overheads related to shared memory, time spent in message passing routines, calling tree information, and performance measures. We will demonstrate the benefits of the data collection mechanism of the MPP Apprentice tool and how it guides the user’s identification of performance problems using application examples from benchmarks and industry.