The Community for Technology Leaders
Green Image
<p>The architecture of parallel machines influences the structure of parallel programs, and vice versa. This symbiotic cycle has been in motion for at least a decade. An important result of research on shared memory applications is the two sets of benchmarks, Splash and NAS, that have driven much research into shared memory architectures and cache-coherence protocols. At the same time, succeeding generations of shared memory multiprocessors (Stanford Dash, MIT Alewife, and Wisconsin Typhoon) have been used to characterize the behavior of shared memory applications. New applications and architectural mechanisms are now emerging. In particular, fine-grain applications are an important emerging class that warrants further study. These applications have long been thought to favor message-passing over shared memory architectures because of their frequent communications and sensitivity to memory latency. The authors present the performance of 14 applications on the Alewife machine, including both coarse- and fine-grain applications. Not surprisingly, Alewife's mechanisms support the good performance of traditional coarse-grain applications from the Splash and NAS benchmark suites. But the authors also show that Alewife provides an excellent communication mechanism for fine-grain applications. The results confirm that hardware support for limited sharing is adequate for a broad range of applications, even on large numbers of processors. Local cache-miss behavior turns out to be important on multiprocessors with low remote miss latencies. The results show that low-latency miss-handling mechanisms for both local and remote accesses in Alewife make fine-grain applications viable candidates for shared memory parallel processing. </p>

B. Lim, F. T. Chong, A. Agarwal, J. Kubiatowicz and R. Bianchini, "Application Performance on the MIT Alewife Machine," in Computer, vol. 29, no. , pp. 57-64, 1996.
88 ms
(Ver 3.3 (11022016))