Return to the Table of ContentsObjective: To establish a testbed for early evaluation of parallel architectures responsive to the computational demands of the Lewis and industry propulsion codes.
Approach: A localized cluster of high-end IBM workstations has been assembled and configured to provide for both distributed memory, MIMD parallel processing and distributed processing applications. Inter-node traffic can be carried via ethernet or one of four high-speed interface options: a low-latency crossbar switch, called ALLNODE; Asynchronous Transfer Mode (ATM); Fiber Channel Standard (FCS); and, FDDI. In addition, a 16-node IBM SP1 has been added to utilize and integrate IBMâ•’s MPP environment.
Accomplishment: A highly flexible configuration of clustered IBM Risc systems has been designed, tested and applied to various industrial codes. The 32-node cluster contains 16 IBM Model 590 and 16 Model 560 Risc Systems, each with minimum of 64 MByte memory, 1 GByte disk, and a CPU benchmarked at 30.5 MFlop (LINPAC). Some nodes have expanded memory (4 with 128 MB, 2 with 512 MB). An IBM Model 970 with 6 GByte disk serves as a resource manager. The cluster has an aggregate maximum of approximately 1 GFlop performance.
Each node is interconnected either using a dedicated ethernet or a low-latency crossbar switch (ALLNODE). The ALLNODE switch was tested in several applications and demonstrated significantly reduced latencies (approx. one-fourth to one-eight) and greater bandwidth (by approximately a factor of four) than a dedicated ethernet. Asynchronous Transfer Mode (ATM) hardware was tested and displayed substantial promise for applications benefiting from low latency.
The 16 node IBM SP1 contains 128 MB of memory and 1GB disk per node. It has ATM connectivity to the cluster system, and its system software was upgraded to an SP2 configuration.
The IBM SP1 was used to run several mission-critical simulations of the MSTAGE code as part of the Stall Line Management Program.
Significance: The testbed environment, combining a Risc Cluster and the IBM SP2 allows for early evaluation of the IBM MPP environment which is intended to provide scalable TeraFlop systems by mid-decade. Its configuration also permits the testing of multiple interconnectivity options. In addition, the cluster is well suited to NASA Lewis' multi-disciplinary approach to aeropropulsion simulation. Different modules of the simulation (e.g. inlet, combustor, etc.) can run on different nodes of the cluster, some possibly parallelizable, others potentially requiring nodes with more memory. Status/Plans: Networking configurations including the low-latency ALLNODE switch and ATM have been shown to substantially improve the performance of parallel applications on the cluster. The current network configuration on the cluster/SP1 environment is ideal for characterizing various interprocessor communication technologies. Enhancements to network approaches resulting from this studies may yield communication speeds approximating the current generation of massively parallel processors. A small cluster of HP workstations will be added to explore heterogeneity issues. In addition, we plan to integrate a GigaRouter as an interconnect between various high-speed networks and devices.
Point of Contact:
Return to the Table of Contents