| For many parallel-computing tasks, clusters of workstations provide performance comparable to parallel machines. At the same time, clusters are cheaper and easier to reconfigure. However, for applications to run successfully on clusters, they must be able to tolerate communication latencies larger than in parallel machines, and they should not require message rates and data transfer bandwidths higher than those available from the cluster interconnect.; This thesis addresses a fundamental limit to the improvement of cluster technologies: the achievable latencies and message rates are limited by the performance of the workstation I/O buses which must be used to interface with the cluster interconnect. We propose to overcome the I/O bus limitations by using an extensible communication architecture. This architecture, called the Distributed Virtual Communication Machine (DVCM), requires the cluster workstations to be equipped with intelligent network interface (NI) cards.; The architecture's main feature is that it enables the transfer across the I/O bus, to the NI communication processor (CoProcessor), of selected communication-related functionality, traditionally executed on the host. To improve the performance of cluster applications, their most communication-demanding modules can be implemented as DVCM-extensions and run on the NI CoProcessor. The architecture admits low-cost implementations, as demonstrated by its current ATM-based implementation, which uses only off-the-shelf hardware. The implementation is described and its performance measured. We show that using the NI CoProcessor for nontrivial tasks is beneficial for several important applications. |
