A summary of the various architectures and configurations used when these performance figures were measured follows. Unless otherwise noted, the test programs were compiled with cc -O.
The Convex SPP1000 and SPP1200 consist of SCI-ring connected nodes (160 MB/second). Each SPP1000 node consists of eight 100 MHz HP PA RISC 7100 processors (120 MHz for the SPP1200) with a cross-bar memory interconnect (250 MB/second). The tests were run under SPP-UX 3.0.4.1 and ConvexPVM 3.3.7.1.
The Cray T3D is 3-D-torus multiprocessor using the 150 MHz DEC Alpha processor. Communication channels have a peak rate of 300 MB/second. Tests were performed using MAX 1.2.0.2. A special thanks to Majed Sidani of Cray for running our communication tests on the T3D using PVM. The PVM communication was with pvm_psend and pvm_precv.
The Intel iPSC/860 is Intel's third generation hypercube. Each node has a 40 MHz i860 with 8 KB cache and at least 8 MB of memory. Communication channels have a peak rate of 2.8 MB/second. Tests were performed using NX 3.3.2. The Intel iPSC/2 uses the same communication hardware as the iPSC/860 but uses 16 MHz 80386/7 for computation.
The Intel Delta is a 512-node mesh designed as a prototype for the Intel Paragon family. Each node has a 40 MHz i860 with 8 KB cache and 16 MB of memory. Communication channels have a peak rate of 22 MB/second. Tests were performed using NX 3.3.10.
The Intel Paragon is a mesh-based multiprocessor. Each node has at least two 50 MHz i860XP processors with 16 KB cache and at least 16 MB of memory. One processor is usually dedicated to communications. Communication channels have a peak rate of 175 MB/second. Test were run under OSF 1.0.4 Server 1.3/WW48-02 and SUNMOS 1.6.2 (using NX message passing).
The IBM SP1 is an omega-switch-based multiprocessor using 62.5 MHz RS6000 processors. Communication channels have a peak rate of 40 MB/second. Tests were run using MPL.
The IBM SP2 is an omega-switch-based multiprocessor using 66 MHz RS6000 processors with L2 cache. Communication channels have a peak rate of 40 MB/second. Tests were run using MPI. The MPI communication was with mpi_send and mpi_recv.
The Kendall Square architecture is a shared-memory system based
on a hierarchy of rings using a custom 20 MHz processor.
Shared-memory latency is about 7 s, and bandwidth is about 32 MB/second.
The message-passing performance was measured using Pacific Northwest
Laboratory's tcgmsg library on one ring of a KSR1 running OSF R1.2.2.
The Meiko CS2 uses SPARC processors with 200 Mflop/s vector co-processors. The communication topology is a fat tree with peak bandwidth of 50 MB/second. The MPSC message-passing library was used for the echo tests. Meiko notes that using point-to-point bidirectional channels in the echo test reduces latency from 82 microseconds to 14 microseconds. A special thanks to Jim Cownie of Meiko for running our communication tests.
The Ncube hypercube processors are custom processors with hypercube communication integrated into the chip. The first generation chip ran at 8 MHz, the second generation chip ran at 20 MHz.
The NEC Cenju-3 results are from a 75 MHz VR4400SC MIPS processor with 32 KBytes of primary cache and 1 MByte of secondary cache using MPI under the Cenju Environment Release 1.5d. Communication channels have a peak rate of 40 MB/second through a multistage interconnection network.
The SGI results are from a 90 MHz PowerChallenge using MPI under IRIX 6.1. The SGI is a shared-memory multiprocessor using a 1.2 GB/s bus.
The TMC CM5 is hypertree multiprocessor using 32 MHz SPARC processors with four vector units and 16 MB of memory per node. Communication channels have a peak rate of 20 MB/second. Tests were run using the message passing library CMMD 2.0.