Large-sci.le stochastic linear programs can be efficiently solved by using a blending of classical Benders decomposition and a relatively new technique called importance sampling. The paper demonstrates how such an approach can be effectively implemented on a parallel (Hypercube) multicomputer. Numerical results are presented. provided crucial programming support in early phases of this project. Access to the iPSC/2 hypercube computer was by courtesy of Oakridge National Laboratory. 1 7 Codes

more »

... U (4 MIPS) couplhd with a 80387 (209 Kflops) numeric ,oprocess(;r for floating jnd/or---ial 9 tII1 point acceleration. It has 4 MBytes of local memory. The hypercube (or Cube) is accessed via a host (or System Resource Manager) which is also a 80386-based system with 8 MByte memory and a 140 MByte hard disk. The operating system on the host is the UNIX System V/386 (Release 3.0). The data transfer rate between the System Resource Manager and the Cube has a peak value of 2800 KBytes/sec. Although the nodes are physically connected as the edges of a hypercube, a trade-marked routing network called DIRECT-CONNECT provides essentially uniform communication linkages between all the nodes. The earlier "store and forward" method used in first-generation hypercubes is replaced by a hardware switching system, the Direct-Connect Module (DCR) on each node. Each DCR provides seven full-duplex channels for internodal communication and one for connection to the System Resource Manager or I/O devices. The network uses a special algorithm for messages longer than 100 bytes. It first sends ahead a header message to the destination node. This header sets gates in each DCR on the intermediate nodes to clear a data path for the message. Once communication with the destination node is established with acknowledgment of receipt of the header, the message is sent through at essentially hardware data transfer rates. The implication of this improved technology is that computational efficiency is essentially independent of the problem domain to machine topology mapping. The hypercube can be programmed as an ensemble of processors with an arbitrary communications network in which each node can communicate more or less uniformly with all other nodes. The host machine allows the user to perform the following tasks. -To edit, compile and link host/node programs. -To access and release the cube (or a partition thereof). -To execute the host program. -To start or kill processes on the cube. Operations peculiar to the hypercube are controlled either by UNIX-type commands OF REPORT OF THIS PAGE OF ABSTRACT UNCLASSIFIED SAR *W. S.;rGO1_28O55C,0 S1acare go'-'9 VV.6