Comparative Study of Distributed Resource Management Systems – SGE, LSF, PBS Pro, and LoadLeveler

Distributed Resource Management Systems (D-RMS) control the usage of hard resources, such as CPU cycles, memory, disk space and network bandwidth, in high-performance parallel computing systems. Users request resources by submitting jobs, which could be sequential or parallel. The goal of a D-RMS is to achieve the best utilization of resources and to maximize system throughput by orchestrating the process of assigning the hard resources to users’ jobs. In the past decade, lots of work have been done to survey and study those systems, but most of them are from the viewpoint of users and D-RMS provided functionalities. In this study, we comparatively study current widely-deployed D-RMS from the system aspect by decomposing D-RMS into three subsystems: Job management subsystem, physical resource management subsystem, and scheduling and queuing subsystem. Also the system architecture to organize these three subsystems are discussed in detail. This work contributes to the D-RMS vendor and research in distributed resource management by presenting D-RMS internals for the designer in their future system upgrade and improvement.