Performance Evaluation of Specialized Hardware for Fast Global Operations on Distributed Memory Multicomputers

Workstation cluster multicomputers are increasingly being applied for solving scientific problems that require massive computing power. Parallel Virtual Machine (PVM) is a popular message-passing model used to program these clusters. One of the major performance limiting factors for cluster multicomputers is their inefficiency in performing parallel program operations involving collective communications. These operations include synchronization, global reduction, broadcast/multicast operations and orderly access to shared global variables. Hall has demonstrated that a .secondary network with wide tree topology and centralized coordination processors (COP) could improve the performance of global operations on a variety of distributed architectures [Hall94a]. My hypothesis was that the efficiency of many PVM applications on workstation clusters could be significantly improved by utilizing a COP system for collective communication operations. To test my hypothesis, I interfaced COP system with PVM. The interface software includes a virtual memory-mapped secondary network interface driver, and a function library which allows to use COP system in place of PVM function calls in application programs. My implementation makes it possible to easily port any existing PVM applications to perform fast global operations using the COP system. To evaluate the performance improvements of using a COP system, I measured cost of various PVM global functions, derived the cost of equivalent COP library global functions, and compared the results. To analyze the cost of global operations on overall execution time of applications, I instrumented a complex molecular dynamics PVM application and performed measurements. The measurements were performed for a sample cluster size of 5 and for message sizes up to 16 kilobytes. The comparison of PVM and COP system global operation performance clearly demonstrates that the COP system can speed up a variety of global operations involving small-to-medium sized messages by factors of 5-25. Analysis of the example application for a sample cluster size of 5 show that speedup provided by my global function libraries and the COP system reduces overall execution time for this and similar applications by above 1.5 times. Additionally, the performance improvement seen by applications increases as the cluster size increases, thus providing a scalable solution for performing global operations