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 runtime 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

Abstract. Event tracing is a well-accepted technique for post-mortem performance analysis of parallel applications. The expert tool supports the analysis of large traces by automatically searching them for execution patterns that indicate inefficient behavior. However, the current search algorithm works with independent pattern specifications and ignores the specialization hierarchy existing between them, resulting in a long analysis time caused by repeated matching attempts as well as in replicated code. This article describes an optimized design taking advantage of specialization relationships and leading to a significant runtime improvement as well as to more compact pattern specifications. 1

Event traces provide a very fine grained view on the performance behavior of a parallel application. Based on this view, performance properties which cannot be represented by profiling data can be described in terms of compound runtime events. In many cases, a compound runtime event indicating the existence of a performance property exhibits a quite complex structure. Most of the relationships by which its constituents are interconnected depend on a specific programming model which makes it difficult to capture all such situations by means of one general representation method. To overcome this problem, we present a generic technique for defining programming model specific abstractions allowing a simple description of compound runtime events in the context of that programming model. In addition, we show that this approach can be easily integrated into the APART specification language and propose appropriate language extensions.

FZJ-ZAM-IB-2004-06

this report, we use data collected from interviews with a number of application developers and developers of load balancing systems in order to specify a general group of event types that are meaningful in many application contexts and that can be customized to fit individual needs. Furthermore, we present a concept for developing a flexible visualization tool that can help in making good use of recorded events. Our considerations apply to parallel application programs running on a multiprocessor system with distributed memory, e.g., a cluster of Ethernet-coupled workstations or an IBM SP2 parallel computer. A substantial portion of the results are meaningful for different types of parallel systems (like shared-memory multiprocessors) as well. 2 Overview of Parallel Systems and their Analysis Requirements