The Economics of Large-Memory Computations

We propose, and justify, an economic theory to guide memory system design, operation, and analysis. Our theory treats memory random-access latency, and its cost per installed megabyte, as fundamentals. We introduce incentives in our economic theory, and side-constraints in our analytic model of hierarchical memory, to ensure su cient memory bandwidth and processor speed in any \well-formed" system of a given latency and size. Our theory suggests that computer users should be charged a \rental" cost, proportional to their use of the total capacity in a hierarchical memory system. This rental cost is a natural unit for algorithmic analysis and, we submit, is a rational basis for pricing. We use our theory to compare the cost/performance of various largememory organizations such as PoPCs (\piles of PCs"), NOWs (\networks of workstations"), SMPs (\shared memory multiprocessors"), MPPs (\massively parallel processors"), and even Cray-class vector supercomputers. Suggested conference track: Performance of Parallel and Distributed Computing Systems.