A General Model for Scheduling of Parallel Computations and its Application to Parallel I/O Operations

The performance of parallel computers for many interesting classes of applications is often limited by the speed of data transfers rather than the speed of computation This observation has motivated the argument that the performance of a parallel computer system should be measured in terms of its data transfer rates both within the system and across a network instead of the peak oating point computation rate This data transfer bottleneck is especially apparent in the case of I O data transfers Technology and application trends indicate that this I O bottleneck is likely to become increasingly important in the future and references therein