Efficient collective communication in distributed heterogeneous systems
Journal of Parallel and Distributed Computing
| Hi-index | 0.00 |
Clusters of workstations have emerged as a popular platform for parallel and distributed computing. Commodity high speed networks which are used to connect workstation clusters provide high bandwidth, but also have high latency. SCRAMNet is an extremely low latency replicated non-coherent shared memory network, so far used only for real-time applications. This paper reports our early experiences with using SCRAMNet for cluster computing. We have implemented a user-level zero-copy message passing protocol for SCRAMNet called the BillBoard Protocol (BBP). The one way latency for sending a 4-byte message between two nodes using the BBP is measured to be as low as 7.8 µs. Since SCRAMNet supports hardware level replication of messages, it is possible to implement multicast with almost the same latency as point-to-point communication. Using the BBP, the latency for broadcasting short messages to 4 nodes is measured to be 10.1 µs and the latency for a 4-node barrier is measured to be 37 µs. We have also built an MPI library on top of the BBP which makes use of multicast support from the BBP. Our results demonstrate the potential of SCRAMNet as a high performance interconnect for building scalable workstation clusters supporting message passing.