Cost-Effective Parallel Computing

Many academic papers imply that parallel computing is only worthwhile when applications achieve nearly linear speedup (i.e., execute nearly p times faster on p processors). This note shows that parallel computing is cost-effective whenever speedup exceeds costup--- the parallel system cost divided by uniprocessor cost. Furthermore, when applications have large memory requirements (e.g., 512 megabytes), the costup---and hence speedup necessary to be cost-effective---can be much less than linear. Introduction Suppose that you need to run many simulations that require large amounts of memory. You may run the simulations on a uniprocessor or a p-processor parallel system. You know that your simulations cannot be parallelized perfectly, so speedups will be less than linear. Parallel simulation will reduce response time, but your task is to maximize job throughput per unit cost, or equivalently, to minimize cost-performance (cost divided by performance). Which system do you select? Conven..