A Performance Monitoring System For Networked Parallel Computing
Phyllis E. Crandall, Nenad Nedeljković, Michael J. Quinn · 1993
We present a graphical performance monitoring system for parallel processing in the workstation network environment. Clusters of workstations have become attractive candidates for parallel processing due to their improving speeds and wide availability. However, network parallel processing introduces new challenges that must be met before this computational paradigm if of practical use. Typically, the connections are Ethernet or Token Ring, which impose bandwidth limitations and message-passing overhead that may negate a large portion of the potential speedup otherwise attainable. The nodes in a workstation network are often heterogeneous: that is, they may have different computational speeds. The combination of network limitations and heterogeneity demands scrupulous partitioning of the workload to minimize communication and maximize computation efficiency. Our system addresses four main issues: 1. Static load balancing, 2. Dynamic load balancing, 3. Runtime analysis, and 4. Graphical ...