High Performance Parallel Interface Research Articles

Supercomputer SX-4 multinode system

The NEC supercomputer SX-4 multinode system series consists of two models, one being HIPPI (High Performance Parallel Interface)-connected model and the other IXS (Internode Crossbar Switch)-connected model. With the IXS, a proprietary high-speed crossbar switch, the HPC (High Performance Computing) up to 1 TFLOPS (Tera Flops) has been enabled by providing the most comprehensive environment for distributed parallel processing. This also means the world's first implementation of a clustered parallel processing. In this paper, we describe the functions of IXS hardware, the new operating system functions, MPI/SX the MPI (Message Passing Interface) processor and NQS/MPI which supports the close cooperation between NQS (Network Queuing System) batch processing system and MPI.

High-Performance Computing Over Switch-Based High-Speed Local Networks

Recent progress in switch-based high-speed Local Area Networks (LANs) makes distributed network computing promising. Three evolving switch-based high-speed networks are the High Performance Parallel Interface (HIPPI), Fiber Channel (FC), and Asynchro

Differential service in high-speed interconnection systems: an algorithm for prioritized access

The requirement to connect high-performance components of supercomputer systems has caused an increasing demand for connection-oriented, high-speed communication in a peer-to-peer environment. The importance of this type of communication is underscored by the advent of the high-performance parallel interface (HIPPI) standard, designed to support the commercial development of interconnection systems for this environment. For such interconnection systems, we present a connection management algorithm that supports priorities to provide preferential access for important classes of traffic, and that assures equitable access for all connections within the same priority class. At the same time, the algorithm utilizes the connection resources efficiently and assures low overhead at the nodes. We demonstrate the algorithm's applicability by outlining a centralized as well as a distributed implementation in an HIPPI-based interconnection system. Furthermore, we introduce an analytic priority model that can be used to evaluate an interconnection system's performance under the algorithm. Using this model, we analyze the efficacy of the nonpreemptive priority discipline and show the load/throughput behavior of a typical system.

Fiber channel workstations interconnection through ATM local network

The Fibre Channel Standard (FCS) is an emerging standard for providing a common efficient transport vehicle for existing channel protocols like the Intelligent Peripheral Interface (IPI), Small Computer System Interface (SCSI) and High Performance Parallel Interface (HIPPI). This paper deals with the interconnection issue of FCS-based workstations through an ATM switch. In particular, the design of a high throughput FCS/ATM InterWorking Unit (IWU) is presented. A specialized segmenting and reassembly protocol, named FAP, has been defined to guarantee the transparency of the end-to-end FCS protocols. The hardware and software characteristics of the IWU are described and its performance evaluated by a simulation study. Performance evaluation has focused on FCS flow control functions with the twofold aim to determine the impact of IWU on the protocol parameters (e.g. timeouts, window size, etc.) and to assume reliable data for a suitable dimensioning criteria for IWU elements (e.g. segmenting and reassembly buffers). Results obtained show that throughput values approaching the available ATM link bandwidth can be achieved depending on the discipline utilized to manage the IWU buffers.

The Rainbow-II gigabit optical network

This paper describes the Rainbow-II optical metropolitan area network (MAN), which supports 32 nodes each at 1 Gbit/s over a distance of 10-20 km. Rainbow-II uses optical wavelength-division multiplexing (WDM), in a broadcast star architecture. Each node uses a separate fixed wavelength for transmitting data and a tunable receiver for receiving one of several data streams. The network is implemented in the form of optical network nodes, each attached to a host computer via the high-performance parallel interface (HIPPI). Each network node contains protocol processing hardware to offload the protocol processing work from the host computer onto the node. The goal is to provide full gigabit-per-second bandwidth to end-user supercomputer applications. Preliminary protocol performance measurements in a testbed network are given.

Architecture and performance of HIPPI-ATM-SONET terminal adapters

The flexibility of the asynchronous transfer mode (ATM) technique to integrate diverse traffic using a single network platform, and the scalability of the synchronous optical network (SONET) to support various transmission speeds up to gigabit per second ranges, have contributed to the adoption of these standards in the future public telecommunications network. Presently, the networking of supercomputers and high performance workstations is widely based on the high performance parallel interface (HIPPI) standard. A number of efforts to allow the access of HIPPI networks to the public ATM/SONET network has resulted in different experimental Gigabit Testbeds such as CASA and Nectar. To provide the access of HIPPI-based local area networks (HIPPI-LAN) to the ATM/SONET network, HIPPI/ATM/SONET (HAS) terminal adapters could be deployed in instances where the cost of evolving the HIPPI network into a full ATM/SONET-based network is higher than that of a HAS interface. This article presents the architecture and performance of a 2.5 Gb/s HAS experimental prototype, which is a key component of the Nectar Gigabit Testbed. It identifies HAS implementation issues and presents possible solutions that can be extended to the implementation of future HAS devices. The HAS architecture is presented, as well as a discussion on flow control implications due to terminating the HIPPI protocol. The HAS performance is then shown through theoretical modeling and measured results.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

Enhanced PVM communications over a high-speed LAN

Performance results of PVM over a local ATM network show the availability of much greater communication bandwidth over traditional LANs such as Ethernet. Application-level performance, however, still lags far behind the capabilities of the physical medium. Realizing the full potential of high-speed networks, therefore, will require further improvements in both hardware and software components of network I/O subsystems.Emulating a parallel machine via a collection of heterogeneous, independent hosts and a general-purpose local-area network has obvious advantages, including cost-effectiveness and very large aggregate processing power and memory. However, the ability of most general-purpose LANs to support communication-intensive parallel applications is debatable. Today, with the emergence of several high-speed, switch-based networks, such as High-Performance Parallel Interface (Hippi), Fibre Channel, and Asynchronous Transfer Mode (ATM), networks that effectively support communication-intensive parallel applications may soon become a reality.Network-based computing offers several advantages. First, independent, commercially available systems and a general LAN can readily incorporate advances in processor and network technology. Second, due to the large memory and processing power available in the aggregate collection of individual host systems, very large applications can execute on a collection of relatively low-priced host systems. Third, the underlying network can support high-speed input/output to applications, for instance, by using disk arrays.One factor that previously fueled much skepticism about the feasibility of network-based parallel computing was the limitations imposed by using traditional LANs, such as Ethernet, as the system interconnect. For many typical network applications that require only occasional file transfers, or infrequent small amounts of data to be transmitted between workstations, an Ethernet-based cluster of workstations will suffice. However, for network-based applications, such as communication-intensive, course-grain parallel applications, traditional networks such as Ethernet simply cannot provide adequate performance. Thus, for this study, we chose a high-speed transport mode as the supporting communication medium.

Impact of gigabit network research on scientific visualization

Networks based on the High Performance Parallel Interface (HIPPI) will become the norm at Los Alamos National Laboratory. The ramification of such a high-speed networking paradigm on scientific visualization are enormous. Not only will scientists have the capability of networked framebuffer animation loops in their offices, but also the partitioning of graphics tasks between MIMD, SIMD, and specialized hardware will be feasible. Of course, as bandwidth increases, the problem size quickly grows to exceed whatever the limits are. For this reason, the investigation of gigabyte networks is currently underway at Los Alamos National Laboratory.

Particle simulation on heterogeneous distributed supercomputers

AbstractWe describe the implementation and performance of a three dimensional particle simulation distributed between a Thinking Machines CM‐2 and a Cray Y‐MP. These are connected by a combination of two high‐speed networks: a high‐performance parallel Interface (HIPPI) and an optical network (UltraNct). This is the first application to use this configuration at NASA Ames Research Center. We describe our experience Implementing and using the application and report the results of several timing measurements. We show that the distribution of applications across disparate supcrcomputing platforms Is feasible and has reasonable performance. In addition, several practical aspects of the computing environment arc discussed.

HIPPI: simplicity yields success

The high-performance parallel interface (HIPPI), a simplex point-to-point interface for transferring data at peak data rates of 800 or 1600 Mb/s over distances of up to 25 m, is reviewed. The HIPPI physical layer, framing protocol, link encapsulation, its mapping to fiber channels, mapping to IPI-3, switch control, fiber extender, and tester are described. The role of HIPPI in networks is discussed.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

HIPPI developments for CERN experiments

The high-performance parallel interface (HIPPI) is an ANSI standard, using a minimal protocol and providing 100 Mbyte/s transfers over distances up to 25 m. A brief technical introduction to the HIPPI is given, followed by examples of planned applications in high-energy physics experiments including the developments involving CERN: a detector emulator, a RISC-processor-based VME connection, a long-distance fiber optics connection, and a HIPPI testbox. Equipment using this standard is offered by a growing number of computer manufacturers. A commercially available HIPPI chip set allows low-cost implementations.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

Network supercomputing

The High Performance Parallel Interface (HIPPI), American National Standards Institute (ANSI) draft standard X3T9, which is likely to become the backbone for interconnecting high-performance computing components into a single unit, is reviewed. The current HIPPI operating environment is described, and preliminary performance measurements obtained while transmitting data between Cray-2 and CM-2 computer systems are reported. The use of HIPPI for computationally intensive applications to reduce their total execution time is discussed. >

Gigabit networking

Factors driving gigabit networking are identified. The problem of wedding the local area networks (LANs) used to interconnect supercomputers to the long-distance networks used for telecommunications is discussed. Changes in computer and telecommunications networks to accommodate higher speeds, longer distances, and new technology are examined. The American National Standard High-Performance Parallel Interface (HIPPI) is considered. Connecting HIPPI to SONET and to FDDI is addressed. Circuit-switched, ring, and tree architectures for advanced networks are discussed.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

The infrastructure for gigabit computer networks

The key features of the synchronous optical network (SONET) and asynchronous transfer mode (ATM) standards for the broadband integrated services digital network (B-ISDN) public network, which is expected to provide a powerful and ubiquitous infrastructure to support the emerging gigabit-per-second data and multimedia applications, are reviewed. The advantages of ATM, the ATM cell format, the B-ISDN protocol reference model and the ATM layer, the ATM adaptation layer, and ATM traffic management and signaling are discussed. An experimental high-performance parallel interface (HIPPI)-ATM-SONET interface is presented as an example to illustrate how the SONET and ATM techniques can be used to provide end-to-end transport for applications requiring a bandwidth of 800 Mb/s and possibly 1600 Mb/s.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>

Control policies for interconected distributed systems via an HIPPI switch

The increase in high-speed communication between supercomputers and between intelligent components of supercomputing systems has lead to the development of the High Performance Parallel Interface (HIPPI) as a standard interface for local high-speed interconnection systems. This interface facilitates connection oriented, one way communication at speeds up to 1600 Mbps. While the standard specifies the physical interface to the interconnection system, it leaves open many aspects of the system, such as switch topology and connection management policy. This paper proposes three different policies for managing the connections in a HIPPI system, a centralized policy, a broadcast policy, and a distributed policy. Assuming a crossbar topology for the switch a analytic model is developed to evaluate the performance of HIPPI interconnection systems under the three connection management policies. The model results can aid in determining the proper size of the interconnection system, and in selecting an appropriate connection management policy.