Network Interface for Message-Passing Parallel Computation on a Workstation Cluster

• Main Author: James C. Hoe
• Other Authors: The Pennsylvania State University CiteSeerX Archives
• Format: Text
• Language: English
• Subjects: Arctic
• Online Access: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.69.9348; http://www.ece.cmu.edu/~jhoe/distribution/mit/hotint94.pdf

As commercial microprocessors become increasingly popular in current MPP architectures, high-performance commercial workstations have also received increased attention as cost-effective building blocks for large parallel-processing systems. The Fast User-level Network (FUNet) project [10] is an attempt at constructing an inexpensive workstation-based parallel system capable of supporting e cient execution of message-passing parallel programs. Based on MIT's Arctic [1] network technology, FUNet connects stockcon gured commodity workstations with a high-bandwidth packet-switched routing network. The Fast User-level Network Interface (FUNi) is the custom hardware network interface device that provides access to FUNet for both message passing and remote direct-memory-access (DMA) block transfers between parallel peer processes on FUNet-connected workstations. The FUNi hardware mechanisms allow direct low-overhead user-level accesses to FUNet while maintaining secure and transparent sharing of FUNet among multiple parallel applications. FUNi can be realized as SBus peripheral cards to allow compatibility with a variety of workstation platforms. The relaxed clock speed (25MHz max.) of SBus allows FUNi to be inexpensively implemented using FPGA parts that are synthesized from designs captured in Verilog Hardware Description Language [15]. SBus's Direct Virtual Memory Access (DVMA)[7] also assists FUNi in overcoming the performance limitations imposed by existing workstation designs. Simulation results have shown that FUNet with FUNi, when coupled with latency-hiding software techniques, is e ective in supporting ne-grained parallel processing on a workstation cluster.