Jim Nichols and Mark Claypool
Powerful, low-cost clusters of personal computers, such as Beowulf clusters, have fueled the potential for widespread distributed computation. While these Beowulf clusters typically have software that facilitates development of distributed applications, there is still a need for effective distributed computation that is transparent to the application programmer. The PANTS Application Node Transparency System (PANTS), enables transparent distributed computation. The system employs a fault-tolerant communication architecture based on multicast communication that minimizes load on busy cluster nodes. PANTS runs on any modern distribution of Linux without requiring any kernel modifications. The initial design and implementation of the load balancing algorithm used only a measurement of CPU usage to make distribution decisions. While CPU usage is the typical metric used in load distribution, other system resources such as disk and memory can become loaded and be a performance bottleneck. In this work, we examine PANTS in the context of load distribution algorithms, build new load indices, develop benchmarks, and evaluate performance. We find our load indices can reduce running time of the benchmarks by 1/2 of the time of the original PANTS indices.
Download:
See also:
Mark Claypool and David Finkel. Transparent Process Migration for Distributed Applications in a Beowulf Cluster, In Proceedings of the International Networking Conference (INC), Plymouth, UK, July 2002. Online at: http://www.cs.wpi.edu/~claypool/papers/pants/
James Nichols and Marc Lemaire. Performance Evaluation of Load Load Sharing Policies on a Beowulf Cluster. MQP-MLC-BW01, Spring 2002. (Advisor Mark Claypool) http://www.cs.wpi.edu/~claypool/mqp/pants-load/