COST-EFFECTIVE PARALLEL COMPUTING

Whereas many would argue that parallel computing is only worthwhile wh en applications achieve nearly linear speedup, this article shows that parallel systems can be cost-effective at modest speedups when memory cost is a significant fraction of system cost. When applications have large memory requirements (like 512 Mbytes), the costup (the parallel system cost divided by uniprocessor cost) can be far less than linear , since parallelizing a job rarely multiplies its memory requirements by p. As a concrete example, the authors use 1994 Silicon Graphics pri ces to show that actual costups can be far less than linear for system s with hundreds of Mbytes of main memory. With real price data for sys tems requiring 512 Mbytes of memory, 8-, 16-, and 32-processor systems are more cost-effective than a uniprocessor when speedups exceed 3.3, 5.0, and 8.6, respectively. This result can be thought of as the conv erse of Amdahl's maxim: Rather than accompanying each 1 MIPS of proces sing power with 1 Mbyte of memory, the authors find that each 1 Mbyte of memory should be accompanied by 1 MIPS of processing power. If one processor does not provide enough power, multiple processors should be used to balance the memory's capacity and bandwidth.