baraglia pSigma: An Infrastructure for Parallel Application Performance Analysis using Symbolic Specifications

In this paper we describe pSigma, which is an in-frastructure for instrumenting parallel applications, en-abling the users to probe into the execution of an appli-cation by intercepting its control-flow at selected points, the selection being made by lucid specifications using symbolic names of data structures and functions in the source program and performance-related information about the memory subsystem. This infrastructure assists application developers and system architects in exam-ining performance events and relating them to the data and control components at the source program level. By providing details on how data is shared among parallel program components, it helps in exposing the perfor-mance effects induced by mutual interference or coop-eration. Our event-based selection schema using sym-bolic names from the source domain results in a very user-friendly instrumentation mechanism.