Newsgroups: comp.parallel From: fils@iastate.edu (Douglas R Fils) Subject: Re: IBM Power Parallel Apparently-To: comp-parallel@beaver.cs.washington.edu Organization: Iowa State University, Ames, IA Follow-Up: comp.parallel Date: Wed, 3 Feb 1993 20:49:47 GMT In article <1993Feb3.164721.23395@hubcap.clemson.edu> CONNOLLY@UKCC.UKY.EDU (John Connolly) writes: > >Does anybody out there have some information on the new Power Parallel >machine announced by IBM yesterday? It seems to be a very reasonable >price for an MPP. I hear about $300K list for a machine rated at at >about a gigaflop (peak). > I got this from the HPCwire site. I removed some of this article that was not related to the question asked. -------------begin included file-------------------------------------- Subject: IBM Upgrades RS/6000 Line, Launches New MPP System (HPCwire Posted: Feb 2 11:49:41 1993 PT) IBM Upgrades RS/6000 Line, Launches New MPP System New York, N.Y. -- Showcasing one of its most conspicuously successful recent ventures, IBM announced a number of extensions and price/performance improvements to its RS/6000 line of RISC workstations and launched its new Model 9076 SP1 Scalable Power Parallel system, based on RS/6000 technology. This was the first formal baptism of the Power Parallel machine, which attracted widespread interest when it was revealed quasi-officially in a peekaboo launch at Supercomputing '92 in Minneapolis last November. A new entry-level desktop workstation, the Powerstation M20, is an extension of IBM's drive into the lower price range long dominated by Sun Microsystems. This color system with 2-D graphics will be priced at $3,995 in a diskless configuration. Other new and enhanced products announced at the same session included an RS/6000-based Network File System (NFS) server, the Model 7051 Power Network Dataserver, priced at $177,600. Developed by IBM's Adstar storage division in San Jose, Calif., the Dataserver is said to be able to accommodate up to 100 or 200 users and to deliver over 2,000 NFS I/O operations per second -- described as three to five times the throughput of other Unix servers. WORKSTATIONS TO MPP The massively parallel (MPP) Unix systems, available initially in configurations from eight to 64 processors, will be "an integral part of the RS/6000 family," according to Irving Wladawsky-Berger, who coordinated the development effort. All of the more than 6,500 applications already ported to the RS/6000 will run on the new parallel systems, he said. The top-end 64-processor system is said to deliver a hypothetical peak performance of 8 GFLOPS. Price/performance for the Power Parallel series will be in roughly the same category as new workstation clusters from Hewlett-Packard, already available in the market, and the Silicon Graphics multiprocessor servers announced last week but not slated for shipment until the first quarter of 1994. Wladawsky-Berger said an entry-level, eight-processor system rated at 1 GFLOPS will be priced at $312,000. A 16-processor version, with abundant active memory and disk capacity and rated at around 2 GFLOPS, will be priced at $600,000 to $700,000. Shipments of the MPP machines will begin in late March/early April with installations at the Argonne National Laboratory, the Cornell Theory Center and Halliburton, a supplier of tools and services to the petroleum industry. An additional 20 or so systems are to be shipped later in the second quarter, and worldwide volume deliveries are expected by the end of the summer. The formal debut date given in an IBM release is October 29. The new systems have appeared only one year after IBM announced its plans to end protracted speculation and uncertainty by throwing its homburg into the MPP ring. The Power Parallel product has been developed by an unusual collaboration among the IBM enterprise division in Kingston, N.Y, the RS/6000 division in Austin, Texas, and IBM's Thomas J. Watson research center in Yorktown Heights, N.Y. Wladawsky-Berger emphasized that marketing will focus intensively upon commercial applications as well as the scientific/engineering market that has, so far, supplied most customers for MPP systems. He said that the architecture is, in principle, scalable indefinitely; cost is the principal practical barrier. As IBM gains additional experience with memory-management and latency issues and performance of the basic chipsets increases further, the goal is to achieve systems with a hundred or so processors that deliver hundreds of GFLOPS or even TFLOPS but are not priced beyond the reach of traditional supercomputer customers. Although existing applications will run on the MPP system, additional adjustments will be needed to take full advantage of parallelism. IBM is working with suppliers of third-party applications for supercomputing to arrange parallelized versions. New software tools, including a Load Leveller tool that enhances management of multiple CPUs, are aimed at assisting users to maximize parallel performance. [...section about price/performance of X-stations deleted...] [...section about PowerPC chip deleted...] -- Norris Parker Smith ----------------------------------------------------------------------------- IBM is a sponsor of HPCwire, helping to make this service available free to users. To learn more about IBM and its products, please see the IBM booth in the Hardware Hall of HPCwire's Trade Show. Copyright 1993 by HPCwire. All rights reserved. HPCwire sponsors - ANS Severen Convex Computer Digital Equipment Federal Data Fujitsu America Genias Software IBM Corp Visual Numerics Intel nCUBE Network Systems MasPar Maximum Strategy MMB Development Corp The Portland Group iOmega Sky Computers The Smaby Group Avalon Hewlett-Packard -- Douglas Fils fils@iastate.edu | Ego vos hortor tantum possum ut amicitiam Dept. of Physics and Astronomy | omnibus rebus humanis anteponatis. Iowa State University (ISU) | Cicero, On Friendship V.17 From: taisuke@taisuke.is.tsukuba.ac.jp (Taisuke -Arai- Boku) Newsgroups: comp.parallel Subject: Re: IBM Power Parallel Date: 4 Feb 93 12:37:51 GMT Reply-To: taisuke@taisuke.is.tsukuba.ac.jp Organization: Inst. of Info. Sci. & Elec., Univ. of Tsukuba, Tsukuba, Japan In-Reply-To: CONNOLLY@UKCC.UKY.EDU's message of 3 Feb 93 15: 03:48 GMT Nntp-Posting-Host: taisuke Follow-Up: comp.parallel Hi, guys In <1993Feb3.164721.23395@hubcap.clemson.edu>, CONNOLLY@UKCC.UKY.EDU (John Connolly) writes: >Does anybody out there have some information on the new Power Parallel >machine announced by IBM yesterday? It seems to be a very reasonable >price for an MPP. I hear about $300K list for a machine rated at at >about a gigaflop (peak). All I know about the machine is as follows: The price $300K is for an entry system with eight RS/6000's. In the maximum configuration, there are four 'frames' (or clusters), each of them consists of 16 RS/6000's. So, the total number of processors is 64 in the maximum. The peak speed of this configuration set is 8 GFLOPS. The intra-cluster and inter-cluster communication is supported by a kind of multistage corssbar switch network called HPS (High Performance Switch). HPS is divided into four parts, and each of them supports the intra-cluster communication between processors in a cluster. All divisions of HPS are also connected. The building block of the network is 8x8 bi-directional crossbar switch, and each division consists of eight switches. They form the 4-base omega network (two stages). Entire HPS consists of 32 8x8 crossbar switches. The data communication throughput is 40MB/s at peak, and the point-to-point communication latency is 500ns (not including software overhead). On each RS/6000, AIX operating system is provided, so that multiprocessing function on each node is provided. For parallel programming, an inter-process communication library is provided. There is also a FORTRAN parallelizing pre-processor, FORGE-90. If there is any missunderstanding, please correct it. -- ------------------------ $@KQ(J $@BYM$(J (Taisuke Boku) taisuke@is.tsukuba.ac.jp $@C^GHBg3X(J$@EE;R!&>pJs9)3X7O(J From: rchrd@netcom.com (Richard Friedman) Newsgroups: comp.parallel Subject: Re: IBM Power Parallel Date: 4 Feb 93 03:45:24 GMT Organization: Applied Parallel Rsrch (Berkeley) Apparently-To: comp-parallel@ames.arc.nasa.gov Follow-Up: comp.parallel >In article <1993Feb3.164721.23395@hubcap.clemson.edu> CONNOLLY@UKCC.UKY.EDU (John Connolly) writes: >> >>Does anybody out there have some information on the new Power Parallel >>machine announced by IBM yesterday? It seems to be a very reasonable >>price for an MPP. I hear about $300K list for a machine rated at at >>about a gigaflop (peak). >> ====================================================================== NEWS RELEASE 3 February 1993 Applied Parallel Research, Inc. ====================================================================== Applied Parallel Reserach 550 Main St., Suite I Placerville, CA 95667 Robert Enk, Sales & Marketing (301) 718-3733, Fax: (301) 718-3734 (Bethesda, MD) FOR IMMEDIATE RELEASE: Applied Parallel Research, Inc. (APR) announces the availability of the FORGE 90 Distributed Memory Parallelizer facility and the xHPF77 batch parallelizer for the IBM Scalable POWERparallel Systems -- the 9076 SP1. These products provide a state-of-the-art tool set for the efficient development and conversion of applications to the 9076 SP1 parallel architecture. The FORGE 90 Distributed Memory Parallelizer is an interactive tool that identifies the loops in a program that are candidates for parallelization, the arrays to be distributed across the processors, and any inhibitors to parallelization. Once the user specifies how the arrays are to be distributed, the system automatically generates the parallel program for the 9076 SP1. A Parallel Profiler is available to provide runtime statistics such as CPU performance, communication times and processor wait times. These statistics are crucial for identifying communication bottlenecks and allows the user to interact with the Distributed Memory Parallelizer to modify data decompositions and loop parallelizations. The FORGE 90 xHPF77 batch parallelizer accepts the High Performance Fortran Explicit Data Distribution Directives in a Fortran 77 program, performs consistency checking on the directives, and then generates a SPMD (Single Program, Multiple Data) program with message passing calls for the 9076 SP1. The Parallel Profiler is also available with xHPF77. John Levesque, President of Applied Parallel Research, said, "We are very excited about the introduction of the 9076 SP1. IBM has recognized the importance of suppying advanced programming tools for overcoming the inherent difficulties associated with programming parallel architectures. The timing of the introduction of the 9076 SP1 coincides well with the availability of our new HPF toolset. Users of the HPF batch parallelizer will be amazed at how easy it is to generate an efficient parallel program." The FORGE 90 tools are fully compatible with IBM's XL FORTRAN and 9076 SP1 libraries, and are availble immediately from APR. Information on the products can be obtained by contacting Robert Enk, APR Directory of Sales & Marketing, in our Bethesda office, (301) 718-3733. ------------------------------------------------------------------------ The following is an excerpt from a News Release issued by IBM on Feb 2 ------------------------------------------------------------------------ IBM ANNOUNCES FIRST SCALABLE, PARALLEL SYSTEM - New Generation of RISC System/6000-based Products Introduced NEW YORK, Feb 2 . . . With system management capabilities and a flexible architecture unlike any other high-performance computer, IBM today announced its first scalable, parallel system -- the 9076 SP1. Based on the powerful RISC System/6000 processor, the general purpose supercomputer provides customers with simultaneous parallel and serial, interactive and batch processing capabilites while achieving outstanding price/performance. Harnessing the power of eight to 64 RISC System/6000 processors and integrating the inter-processor communication efficiencies of an innovative switch technology, the new IBM system has been well-received by several early users. The Scalable POWERparallel System 9076 SP1, the initial member of IBM's new Scalable POWERparallel Series of computers, uses the AIX/6000 operating system. The high-performance computer can run a single, parallel job using multiple processor nodes, execute serial tasks on individual nodes, or perform a combination of these tasks simultaneously. This general-purpose nature of the SP1 enables customers to explore and develop parallel applications while concurrently running their day-to-day serial production work within the SP1 AIX/6000 environment. The compute power and architecture of the SP1 is ideally suited for scientists, engineers, researchers and analysts who require greater power to solve complex, numerically-intensive computational problems. These problems can be found in application areas involving financial modeling, long-range weather forcasting, numerical analysis, geophysical modeling, pharmaceutical design, and advanced aircraft and automotive design. PERFORMANCE: Computing power increases with the number of frames included in the system. Based on the number of frames, four major configurations of the SP1 are being offered. These configurations provide customers with modular processing power ranging from 1 to 8 gigaFLOPS. The peak performance per individual processor node is 125 megaFLOPS. A single frame houses eight to 16 processor nodes, the redundant power supply and an optional high performance switch. The SP1 can accommodate up to four frames. HARDWARE: Each RISC System/6000 processor can be configured with a maximum of 256MB of memory and 2GB of disk; a 64-way system contains a maximum memory of 16GB and 128GB of disk. A full version of AIX/6000 will run on each processor node. System management and maintenance of the SP1 is provided by an IBM RISC System/6000 control workstation, using IBM System Support Programs. Central to the scalability and high-performance features of the SP1 is the new IBM High-Performance Switch which enables high-speed inter-node communication for parallel applications. This advanced technology, based on years of IBM research in parallelism, is designed to support a growing number of processor nodes, well beyond initial configurations. The new IBM AIX Parallel Environment Software provides support for parallel applications devleopment, analysis and execution. It offers a parallel application interface that enables maximum switch performance and tools for parallel debugging and profiling. The software also features innovative program visualizaiton and performance monitoring tools for application analysis. In addition to the IBM AIX Parallel Environment, several parallel application development packages will be available on the SP1, including Express, Linda, and PVM. Existing parallel applications running in these environments can be readily ported to the system. FORGE 90 will also be available to enable High Performance FORTRAN (HPF) parallel programs to run on the SP1 and assist in parallelizing existing serial FORTRAN programs for the system. IBM is currently working with approximately 30 software vendors in developing scientific and engineering applications for the SP1 and future POWERparallel systems. Pricing for the SP1 is dependent upon the type of configuration and customer-specific requirements. An entry model, 8-processor system is priced at $312,000. Built-to-order systems can be requested immediately. Volume shipment of the SP1 is scheduled for 3Q93. ------------------------------------------------------------------------ /// Applied /// FORGE 90 Customer Support Group /// Parallel /// 550 Main St., Placerville, CA 95667 /// Research, Inc. (916) 621-1600 621-0593fax forge@netcom.com ------------------------------------------------------------------------ -- /\=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=/\ \/Richard Friedman (510)528-7055 | rchrd@netcom.com \/ /\Applied Parallel Research (Berkeley)| /\ \/=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=-=\/