An adaptive cost system for parallel program instrumentation

Abstract

We present a new data collection cost system that provides programmers with feedback about the impact data collection is having on their application. We allow programmers to define the level of perturbation their application can tolerate and then we regulate the amount of instrumentation to ensure that threshold is not exceeded. Our approach is unique in that the type of data gathered remains constant; instead we regulate when it is collected. This permits programmers to trade speed of isolation of a performance problem for less application perturbation. We implemented this cost system in the Paradyn Performance Tools and present case studies demonstrating the accuracy of the cost system.