Narrow Results

In this paper, we report the recent progress in parallel processing research in Japan.

The demand for high-performance supercomputers has been growing rapidly in the scientific and engineering fields. To meet this demand, Fujitsu has developed new pipelined supercomputers, the VP2000 series, which attain a maximum performance of up to 5 GFLOPS. This paper describes the VP2000 series in the following areas: architecture, hardware features, hardware technology, multiprocessor system, and software features.

The Quantum Flux Parametron (QFP) is a Josephson Junction device using the polarity of a quantum magnetic flux to represent a binary information. A quantum unit of magnetic flux in a QFP can be controlled to realize majority logic operation, consequently AND and OR logics can be realized. Together with negation, which can be realized by a flux-inverting transformer, various computing logic devices can be constructed. In addition, the QFP is also a latching device so that a flux can be confined in the device to realize memory function. Due to the homogenous nature of memory and logic circuits as well as the flux nature of the signal which can be transmitted inductively, an entirely new Josephson supercomputer with memory and CPU pipelining arranged in a 3D setting was conceived. The computer with 40 processors is expected to run at a maximum speed of 20 GFLOPS of scalar processing.

Since the validity of the Navier-Stokes equations is well established, any fluid dynamic phenomenon could be calculated if methods for solving them correctly are obtained. A fair portion of this dream seems to have come true through the remarkable development of supercomputers and solution algorithms that have made the simulation of high-Reynolds-number flows possible. For understanding the underlying flow mechanism, means of properly visualizing the computed flow field are needed and have been developed. On the whole, computer simulation is becoming the most effective tool for the study of fluid dynamics.

Operational numerical weather prediction requires that all available main-frame computing power be harnessed to the one task, in order to maintain operational schedules. ECMWF therefore became involved in parallel processing as soon as it became available on supercomputers. The main operational programs were restructured to use multitasking on a limited number of processors, currently 8.

The availability of supercomputers with massively parallel architecture which is expected in the near future raises quite a number of major problems. These areas have been discussed in the biennial workshops held at ECMWF since 1984. Besides the user environment, the inter-dependency of hardware architecture and algorithms needs most attention. To study these questions, ECMWF became involved in the GENESIS project sponsored by the European Community. Initial results of this co-operation in the form of benchmark results will be presented.

The challenges posed by new developments in weather prediction strategies will be highlighted and an assessment of the impact parallel computing will have on their being met will be made.

Multidimensional torus networking topology has become widespread recently in the domain of high-performance computers, clusters, and grids, as well as in the domain of networks on chip. Torus represents an ideal communication structure with the shortest distance and multitude of alternative shortest paths between a pair of nodes. We study simple and powerful local packet forwarding (switching) rules that provide packet delivery with quasi-optimal load balancing and do not use tables of addresses (routes). Implementation of these rules in the form of micro-program code within switching nodes increases considerably network performance, security, and QoS. We use infinite Petri nets and reenterable models in the form of colored Petri nets for prototyping the multi-dimensional torus interconnect simulator to study and compare the packet forwarding rules. Then, an ad-hoc simulator of torus interconnect ts is implemented in the C language to provide high performance and the possibility of simulation over prolonged intervals of time. The simulation results acknowledge the advantages of local packet forwarding rules.

In this paper, we extend an information-theoretic approach of computer performance evaluation to supercomputers. This approach is based on the notion of computer Capacity which can be estimated relying solely on the description of computer architecture. We describe the method of calculating Computer Capacity for supercomputers including the influence of the architecture of communication network. The suggested approach is applied to estimate the performance of three of the top 10 supercomputers (according to TOP500 June-2016 list) which are based on Haswell processors. For greater objectivity of results, we compared them relatively to values of another supercomputer which is based an Ivy Bridge processors (this microarchitecture differs from Haswell). The obtained results are compared with values of TOP500 LINPACK benchmark and theoretical peak and we arrive at conclusions about the applicability of the presented theoretical approach (nonexperimental) for performance evaluation of real supercomputers. In particular, it means that the estimations of the computer capacity can be used at the design stage of the development of supercomputers.