Introduction

With the development of new generations of large-scale parallel machines comes the question of what performance data is important to observe and how it should be observed. While the increasing complexity of high-performance systems suggests that new machine features will require performance measurement facilities different from existing tools, there is a strong need to create empirical performance observation technologies that are cross-platform, reusable, and open source. Robust performance technology must address the dual goals of portability and specificity, providing for the latter means extending existing observation capabilities to support machine-specific features, without sacrificing usability and efficiency in how the technology is applied.

In this paper we discuss how the TAU parallel performance system contends with issues of system diversity while maintaining a consistent performance measurement model and observation flexibility. The parallel machine environment we study is the Cray XT3, in particular, the system interfaces for extracting performance and memory information. We describe how TAU incorporates these interfaces in the general TAU measurement facility. In Section §2 , we describe in detail TAU's instrumentation approach and capabilities. TAU's measurement options are explained in Section §3. Section §4 shows how we have enabled TAU to use the memory introspection capabilities in the XT3 Catamount compute node kernel. Results from experiments are shown. Conclusions are given in Seciton §5.