 |
| Home Alewife AlphaServer SC AP 3000 AV 25000 The Challenge CM-2 CM-5 Dash E10000 Exemplar V21000 Exemplar V2600 Flash Gamma II Plus Hector ILLIAC IV iPSC KSR-2 MP-1216 NCUBE/ten NUMAchine Paragon SR8000 SX-5 T3E VPP5000 |
| The IBM RS/6000 SP Machine type: RISC-based distributed-memory multi-processor cluster Models: IBM RS/6000 SP. Operating system: AIX (IBMs Unix variant) Connection structure: -switch Compilers: XL Fortran (Fortran 90), HPF, XL C, C++ |
| In the present SP systems 5 types of nodes can be used: 2 types based on the 332 MHz PowerPC 604e CPU: thin and wide nodes that differ in the number of PCI slots available on the node. As the PowerPC CPU can deliver 2 floating-point operations/cycle, the peak speed per processor is 664 Mflop/s. The nodes can have either 2 or 4 CPUs addressing a common memory. Furthermore, there are 3 POWER3-II based nodes: thin, wide and high. The first two variants again only differ in the amount of PCI slots (2 against 10). Both can harbour 2 or 4 POWER3 CPUs that draw on a shared memory thus making it SMP nodes as in the PowerPC-based nodes. The high nodes can accommodate up to 16 POWER3s and 5 PCI slots. Again these 16 CPUs share the memory in the node and is an SMP processor on which programs can be parallelised using for instance OpenMP. IBM positions the POWER3 SMP node systems primarily for the technical/scientific market. In the parameter list above we included the presently fastest POWER3 processors. And only considered the ``high node'' configuration. The SP configurations are housed in columns, "tall" or "short" frames, of which the tall frames can contain 8--16 processor nodes and short frames half of the tall frames. How many actually are installed depends on the type of node employed: a thin node occupies half of the space of a wide or high node. Each frame is recommended to be configured with at least one wide node, although a frame completely filled with thin nodes seems possible according to the documentation. The RS/6000 SP is accessed through a front-end control workstation that also monitors system failures. Failing nodes can be taken off line and exchanged without interrupting service. The so-called high-performance switch that connects the nodes is an -switch as as described in the section on SM-MIMD systems and, although we mentioned only the highest speed option for the communication, the high-performance switch, there is a wide range of other options that could be chosen instead: Ethernet, Token Ring, FDDI, etc., are all possible. The high performance switch is the third generation of this interconnect. The single-direction bandwidth is quoted as 500 MB/s and tests with MPI-based point-to-point communication from the EuroBen Distributed memory benchmark have shown that one can come very close to this limit. The PowerPC-based models are not able to use this third generation switch so the switch connection them has a nominal bandwidth of 200 MB/s of which a bandwidth of about 160 MB/s can be attained using MPI. Applications can be run using PVM or MPI. Also High Performance Fortran is supported, both a proprietary version and a compiler from the Portland Group. IBM uses its own PVM version from which the data format converter XDR has been stripped. This results in a lower overhead at the cost of generality. Also the MPI implementation, MPI-F, is optimised for the RS/6000 SP systems. The standard commercial models that are marketed contain up to 128 nodes. However, on special request systems with up to 512 nodes can be built. In fact, the POWER3 SMP node is a first commercial spin-off of the ASCI Blue Pacific system with more than 1300 processors. Measured Performances: A performance a performance of 7226 Gflop/s for an 8000 processor system is reported for solving a 518,096-order dense linear system yielding an efficiency of 60%. In speeds of 1.2 and 1.0 Gflop/s are quoted for solving a full linear system and a full symmetric eigenvalue problem of order 1000, respectively on a single CPU. |
| System Parameters: |
| Model SP PowerPC 604e SP POWER3-II Clock cycle 332 MHz 375 MHz Theor. peak performance Per Proc. (64-bits) 664 Mflop/s 1.5 Gflop/s Per 4-proc. SMP node 2.66 Gflop/s 6.00 Gflop/s Per 16-proc. SMP high node --- 24.0 Gflop/s Maximal 340 Gflop/s 3.07 Tflop/s Main memory Memory/node <= 16 GB <= 64 GB Memory/maximal <= 1 TB <= 4 TB No. of processors 8-2048 8-2048 Communication bandwidth Point-to-point 200 MB/s 500 MB/s |