Stealing cycles: Can we get along?

The exploitation of networked workstations as a set of under-util ized computational resources is an attractive idea for cost effective parallel computing. The use of these resources, however, raises a number of interesting policy and performance questions (e.g, what impact will a parallel application have on the normal workstation user, what performance levels can be expected when many users are logged on?). We present performance models for parallel computations executing on shared, networked workstations that can help to address these questions.