StarT - Jr : A Parallel System from

StarT-jr is an experimental parallel system composed of a network of personal computers (PCs). The system leverages the momentum of the microprocessor and PC industries to achieve excellent single node performance at a low cost. For parallel processing, StarT-jr uses the Flexible User-level Network Interface (FUNi) to provide low-overhead, user-level inter-processor communication over two IEEE 1394 High Performance Serial Busses. This eecient message-passing mechanism enables StarT-jr to exploit ne-grained parallelism for good parallel performance. FUNi is based on an embedded processing system on a PCI card. Custom network hardware assembled from a commercial IEEE 1394 chip set provides FUNi with access to the IEEE 1394 network. In message passing, FUNi's embedded processor serves as a network coprocessor and manages an user-accessible message-passing interface in the host memory. User-level applications directly manipulate the interface location in host memory using cached reads and writes. Costly physical I/O accesses to device registers on the PCI bus are avoided. Currently, FUNi can eeciently support both ne-grain message passing and direct memory-to-memory transfers of large data blocks. FUNi can also support globally coherent shared memory by capturing and responding to memory accesses within a designated global address range. FUNi maintains a globally coherent shared memory cache to minimize global memory access latency. The necessary coherence protocol processing and communication is performed by the FUNi coprocessor. We have demonstrated a two-node prototype of StarT-jr and are awaiting fabrication of additional interface cards in order to assemble an eight-node system. StarT-jr currently supports an active message-based lightweight communication library for the C programming language. Preliminary measurements of the communication library demonstrated overheads of 14 sec for sending or receiving small (40 bytes) messages, and an user-to-user latency of 85 sec. Direct memory-to-memory transfers can sustain 3.4 MByte/sec on an unloaded network. With regard to the shared memory operation, reading a shared-memory location cached in FUNi takes approximately 2 sec.