I/O»Aware Gang Scheduling

Out-of-core applications are becoming increasingly prevalent and important. As the number of processors allocated to such applications increases, a greater percentage of an application's working set can be stored in memory, which usually results in fewer disk accesses. This feature can greatly improve the application's execution time beyond the usual speedup from increased parallelism. In fact, when enough processors are assigned to an application to make it fit in core, a super-linear speedup can be achieved. Most gang schedulers are oblivious to the I/O requirements of an out-of-core application. We propose modifying gang schedulers to factor an application's I/O behavior into their allocation decision. Given this information, we investigate the benefit of reducing or minimizing I/O accesses to disk and compare this strategy with current gang scheduling algorithms. Our results show that I/O-aware gang schedulers can dramatically reduce an application's turnaround time and increase throughput on a multicomputer. When compared with conservative processor allocation algorithms that underestimate the processor requirements, as much as a 11.6 fold improvement in average turnaround time resulted. When compared to algorithms that assign all the processors to the applications, there was as much as a 35% improvement.