KOJAK - A tool set for automatic performance analysis of parallel programs

Abstract

Today's parallel computers with SMP nodes provide both multithreading and message passing as their modes of parallel execution. As a consequence, performance analysis and optimization becomes more difficult and creates a need for advanced performance tools that are custom made for this class of computing environments. Current state-of-the-art tools provide valuable assistance in analyzing the performance of MPI and OpenMP programs by visualizing the run-time behavior and calculating statistics over the performance data. However, the developer of parallel programs is still required to filter out relevant parts from a huge amount of low-level information shown in numerous displays and map that information onto program abstractions without tool support. The KOJAK project (Kit for Objective Judgement and Knowledge-based Detection of Performance Bottlenecks) is aiming at the development of a generic automatic performance analysis environment for parallel programs. Performance problems are specified in terms of execution patterns that represent situations of inefficient behavior. These patterns are input for an analysis process that recognizes and quantifies the inefficient behavior in event traces. Mechanisms that hide the complex relationships within event pattern specifications allow a simple description of complex inefficient behavior on a high level of abstraction. The analysis process transforms the event traces into a three-dimensional representation of performance behavior. The first dimension is the kind of behavior. The second dimension describes the behavior's source-code location and the execution phase during which it occurs. Finally, the third dimension gives information on the distribution of performance losses across different processes or threads. The hierarchical organization of each dimension enables the investigation of performance behavior on varying levels of granularity. Each point of the representation is uniformly mapped onto the corresponding fraction of execution time, allowing the convenient correlation of different behavior using only a single view. In addition, the set of predefined performance problems can be extended to meet individual (e.g., application-specific) needs. © Springer-Verlag 2003.