On Increasing Resource Utilization of a High-Performance Computing System

Abstract

In a render farm, a render host usually requires high I/O bandwidth for reading scene files and writing rendering results. A multi-core render host may only activate partial cores in order to restrict its total throughput to the bandwidth limitations. This paper investigates feasible ways for using resource fragments formed by inactivated cores. We develop a parallel application consisting of single-core malleable tasks with small input/output data, and demonstrate that it can be scheduled to adaptive to the unstable resource fragments of the render farm. We develop a broker to collect idle processor cores and control the execution of the malleable tasks in background mode. The experimental results show a significantly increasing on resource utilization of the render farm without interfering with the original rendering tasks.