Using a Hypercube Algorithm for Broadcasting in Internet-Based Clusters

Internet-based clusters of workstations have been extensively used to execute parallel applications. Although these internet-based clusters seem to be an easy and inexpensive way of obtaining great performance, it may not always be so. When using such a cluster for executing a parallel application, performance may not be as good as expected due to delays in communication. Also, the heterogeneity in communication makes it hard to take advantage, or reuse, communication strategies that were useful in regular-topology platforms, e.g., parallel machines or LAN-based clusters of workstations. For instance, broadcasting in an internet-based cluster may be more challenging due to the variety of communication links and, consequently, of point-to-point latencies. In this paper, we present a strategy to improve hypercube-based broadcasting algorithms that are used in regular-topology platforms, so that they can execute efficiently in internet-based clusters of workstations. 1Introduction Recent advances in high-speed networks and improved microprocessor performance are making clusters of workstations an appealing vehicle for cost-effective parallel computing. Clusters of workstations built with commodity hardware and software components are playing a major role in supercomputing [12], being used effectively as parallel machines for large, parallel, scientific applications. With the advent of the Internet, these clusters have been evolving from homogeneous machines connected by a LAN (Local Area Network) to different kinds of machines connected by the Internet. In fact, there are several groups working on systems that enable users to " steal " cycles from idle (or seldom used) machines located in other departments or even other institutions. Some examples are the NOW Project [1] from the University of California, Berkeley and the CONDOR System [6] from the University of Wisconsin. Although these internet-based clusters seem to be an easy and inexpensive way of obtaining great performance, it may not always be so. When using such a cluster for executing a parallel application, performance may not be as good as expected due to delays in communication. Internet-based clusters are connected by heterogeneous links, which may or may not include long-distance high-latency low-bandwidth paths. Depending on the application's communication pattern and computation/communication ratio, the delays imposed by these slow paths may affect or not the performance of the application. The heterogeneity in communication makes it hard to take advantage, or reuse, communication strategies that are useful in regular-topology platforms, such as parallel machines or LAN-based clusters of workstations. For instance, broadcasting in an internet-based cluster may be more challenging …