High-performance computing systems are currently used by many scientists and researchers with Job Management Systems (JMS) that are deployed over high-performance computing systems. To achieve high performance in high-performance computing, systems communication performance must be improved in the interconnect. JMS sometimes adopts the staging, witch is a method to transfer data used by an application program to a local high-speed storage of computing nodes so that the access time to the data is reduced. In this type of high-performance computing systems, the inter-process and the staging communication traffic coexist in an interconnect. In this paper, we investigate the mutual influence between the inter-process and the staging communication traffic and explore a way to exclude the influence of the staging and inter-process communications by targeting a cluster system with a fat-tree interconnect. For this purpose, we propose two types of conflict avoidance methods between the staging communication traffic and the inter-process communication traffic
namely, a link separation conflict avoidance method and a link sharing conflict avoidance method. The link separation conflict avoidance method separates interconnect links into the staging communication traffic and the inter-process communication traffic when the staging execution takes place. The latter proposed method allocates a route to each traffic on the shared interconnect while taking the bandwidth usage into account. Both proposed methods use Software Defined Networking to get the bandwidth usage and to control both types of traffic. We performed an experimental simulation on our deployed virtual cluster system built on a single machine to investigate how each proposed method behaves when the staging communication and the inter-process communication take place in a fat-tree interconnect. In the simulation, the staging time, application execution time and job execution time were measured for each proposed method and for the Equal-Cost Multi-Path method as the existing load distributed method was applied. The result shows that both of the proposed methods improve the performance of the staging and the inter-process communication.
