A Low Cost Method for Determining the Entire Network Path Bandwidths of Indirect High-Performance Computer Networks

摘要

Bandwidth is an important performance metric for high-performance computer network (HINET). How to determine large-scale HINET path bandwidths quickly and precisely is a highly pivotal challenge for parallel application performance analysis and prediction, optimizing end-to-end data transport performance, network management, etc. However, it is hard to determine the entire network point-to-point bandwidths of HINET since it often costs amounts of measurement time. Existing path bandwidth measurement methods focus on improving the accuracy of single path bandwidth measurement, however, lack considerations for reducing the entire network path bandwidths measurement times. To address this issue, we present an efficient method to generate the full-scale point-to-point bandwidth measurement plan for rapidly obtaining each path bandwidth in HINET. For evaluating, a source routing Fat-tree simulator and comprehensive experiments were conducted for different application scenarios. Both experimental and analysis showed that our method can reduce the measurement times significantly. In particular, compared to the exhaustive method, the measurement times of our method are one in 185 for the entire path bandwidth fault and can save 95% measurement times in 1024 nodes network with synthetic traffic.