SHRIMP performance monitor: design and applications

Growing complexity in many current computers makes performance evaluation and characterization both increasingly difficult and increasingly important. For parallel systems, performance characterizations can be especially difficult to obtain, since the hardware is more complex, and the simulation time can be prohibitive. The challenge is to design a low-cost and yet flexible and powerful performance monitoring system to provide systems implementers and application programmers with detailed performance information. This paper describes a performance monitoring system for the SHRIMP multicomputer. The system's core is a hardware monitor with several novel features including multi-dimensional histograms, page tags, histogram categories, and a threshold interrupt mechanism. We also describe software applications that make use of these features. These applications range from fairly simple code-oriented or data-oriented performance tools, to more complicated on-the-fly use of the monitor to improve the performance of a shared virtual memory system. We have found that the concurrent development of the hardware and software portions of the system has led to a novel design that supports a wide range of hardware and software uses.