An evaluation of the integration of reconfigurable hardware with the network interface in cluster computer systems.

摘要

Demand for high performance computing continues to remain strong despite the ever increasing performance of commodity desktop machines. Both Beowulf Clusters and reconfigurable computing provide high performance alternatives. While each technology offers impressive performance to certain applications, each technology also has significant limitations. Unfortunately, a straight-forward integration of the two technologies yields a system that has an even narrower range of application than either technology independently.; This work focuses on incorporating reconfigurable computing in Beowulf Clusters in a cost-effective manner. To achieve cost-effectiveness, the addition of reconfigurable computing to a cluster must benefit a wide range of applications. The commodity components in a Beowulf Cluster prohibit reaching this goal for many applications. This limitation is overcome by combining reconfigurable computing with a high-performance network interface to form an intelligent network interface card (INIC). Placing reconfigurable computing in this data path enhances cluster performance by: providing hardware implementations of collective operations, enhancing the computing capabilities of the cluster nodes, and providing high-performance dataflow processing of network traffic.; Performance models are developed to assess this new architecture for two micro-benchmarks and three applications. These models are evaluated with a prototype INIC for one of the micro-benchmarks and two of the applications. Results indicate that the micro benchmarks are improved by a factor of two and application performance improvements range from 10% to a factor of four. Based on these results, a next-generation card was designed to achieve significantly better performance. A cost analysis was performed for the prototype and next-generation INICs. This analysis indicates that the prototype INIC is cost-effective if the cost of the INIC is constrained. The higher performance of the next-generation design leads to cost-effectiveness over a much larger range of INIC costs.