Colloquium Speaker

Speaker:Professor Clark Thomborson
Computer Science Department
The University of Auckland
Auckland, New Zealand
Topic:The Economics of Large-Memory Computations
Date:Thursday, November 18, 1999
Time:9:30 AM
Place:Gould-Simpson, Room 701

Refreshments will be served in the 7th-floor lobby of Gould-Simpson at 9 :15 AM


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 for latency-bottlenecked computations. For such computations, we define a natural concept of work (prebs, or petareference-bytes) and memory quality (prebs/dollar).

Of course, not all computations are latency-bottlenecked. We will argue, however, that CPU- and bandwidth-bottlenecks are indications of poorly configured hardware or poorly tuned software, for many computational problems of economic importance. In any case, our metric of latency-bottlenecked work is a useful adjunct to the more common metrics of CPU-seconds for CPU-bottlenecked computations, and of channel utilization for bandwidth-bottlenecked systems.

Our economic theory has immediate and near-term applications in diverse areas. It allows us to compare the cost/performance of various large-memory 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. We use it to measure the efficiency of an operating system at supporting work by an application, without excessive overhead in memory transfers. We could provide guidance to performance programmers, by reporting computational progress and efficiency under our metric. Finally, we extend our economic theory into an analytic model of the CPU resources, degree of parallelism, capacity, latency and bandwidth in the various levels of any "well-formed" memory hierarchy, thereby providing guidance to computer architects.