Len Dekker

Architecture and Programmature of the MIMD-Structured Delft Parallel Processor

In this paper a parallel processor with a MIMD-structure, the so-called Delft Parallel Processor (DPP), will be presented. The DPP, which has been developed at Delft University of Technology, is operational since spring 1983. The architecture and the developed software for programming the DPP will be discussed. The last part of this paper will be a reflection on the general thoughts about a hierarchical hardware/software computing system with a distributed MIMD-structure.

Optical link in the Delft parallel processor—an example of MOMI-connection in MIMD-supercomputers

Abstract Much attention is given to the problem how to avoid transfer-bound processing in MIMD supercomputers. The interactivity between processing tasks is compared with the interconnectability that exists between processors in the case of a multi-bus communication system. For this purpose quantitative measures are introduced for both the task interactivity and the processor interconnectability. For a MIMD computer with p processors and p busses the asymptotic speed-up is proportional to p both for tightly and loosely coupled tasks. For tightly coupled tasks a still larger speed-up can be achieved by taking more than p busses. For a MIMD computer with p processors in the case of 2-level parallelization considerably more speed-up can be obtained, but that requires at the 2nd level of parallelization p powerful interconnects (one per processor). Full processor interconnectability is ideal in the sense that no queueing problems can arise. In case of 1-level parallelization with p processors full interconnectability requires p 2 (one-word wide) interconnections. In the Delft Parallel Processor (DPP), instead of using a p 2 -tuple bus communication system, for this purpose a multibroadcast system with p data channels has been applied, where each data channel has p taps (one per processor) and each processor has a p-tuple accessible input memory (one input per channel). In the DPP84 (with maximally 16 processors) electric data channels have been applied. But for technical reasons for large p full interconnectability is only feasible by the way of optical data channels. This optical interconnect must be provided with electro-to-optic (E/O) and opto-to-electric (O/E) transducers as long as optical computing is not yet practically possible. The resulting Electro-Optic Communication System (EOCS) will be implemented in the DPP8X. The EOCS will consist of a combination of guided-wave and a p-tuple way starcoupler technique. Special E/O and O/E transducers have to be developed as well. An intelligent Opto-Electric logic element, the POWERRAM, is realized as a prototype p-tuple accessible input memory IC, capable to accept the multidata stream at the input of a processor in one clock cycle.

Parallel simulation of 3-D flow and transport models within the NOWESP project

Abstract In this paper, a short overview of project 3, “Advanced Flux Modelling”, in the NOWESP project of the MAST II programme is given. The computational requirements of large scale 3-D flow and transport models of the North-West European Shelf are considered. The possibilities and implications of massively parallel processors for the simulation of these models are discussed. We also present some of the results and experience obtained in parallelising the shallow water flow and transport models within the NOWESP project.

Optical interconnects in a multicomputer environment

The Deift Parallel Process, based on a multi-computer architecture and constructed with a Multi-Broadcast Interconnection topology, is a 16 processing no&s Multiple Instruction stream, Multiple Data stream computer developed at the Deift University of Technology in 1977 to evaluate software algorithms, architectures and interconnection topologies experimentally. The Processing Elements in the nodes, containing Weitek Floating Point Multiplier/ALU combinations, are going to be interconnected in a full parallel Electro-Optic Communication System. The Routing Elements, containing Transputers, are linked by an optical reconfigurable network, which enables a simultaneous communication between all the REs.

Current status and future research of the Delft 'supercomputer' project

In despite of the extensive class of computers a break-through of the implementation of Optical Interconnects in the Massive Parallel versions has still not taken place. Only a top-bottom approach will give a thorough insight where those techniques will satisfy forcing researchers in this area to cover almost all the knowledge of topics related to Computer Archi- Lectures Transport media Electro-Optic and Opto-Electronic technologies. Interconnecting Printed Circuit Boards through an Optical Backbone could be the first step© (1991) COPYRIGHT SPIE--The International Society for Optical Engineering. Downloading of the abstract is permitted for personal use only.

Massively parallel processing: optical interconnects according to a system to device approach

Despite the undisputable success in the telecommunication area, applications of optical interconnect techniques to the crate-to-crate, node-to-node, chip-to-chip and gate-to-gate level within a Massively Parallel Computer architecture have still failed to materialize. Questions, such as whether to choose for a serial ultra-high or a parallel moderate data throughput rate, always considering a logic-to-logic approach, are still not adequately answered. An unavoidable number of thresholds a computer manufacturer has to conquer, in case a decision has to be made whether to choose for an electrical or optical solution. An overall system-to-device observation eases in a solution of this dilemma. The authors overview the R & D exertions of the Opto Electronic Systems and Engineering Group of the Delft University of Technology with respect to the development of an Optical Backplane System, enabling a free space data transport in a Massively Parallel Processing environment.<<ETX>>

Simulation of systems

The appearance of the first issue of the journal Simulation Practice and Theory gives the readership an opportunity to become informed about aims and scope of this journal. As Editorin-Chief I will present my belief that there is need of this journal today and in the foreseeable future. I do this by giving my view on the role that simulation of systems plays in the modern world with often a large technological, economic and social impact as tool in experimental studies on systems. Everyone has experience with living systems and has contact with other dynamic systems like e.g. environmental and technical systems. Consequently one has got some notion of a dynamic system within its environment, inputs to a system from its environment, the state of a system, and outputs from a system to its environment. From contacts with systems human beings learn in life to know systems and to understand their behaviour in time and space, where one conceptualizes the knowledge about a system into a conceptual model. Also experimentation on a system giving stimuli to the system and observing the response forms part of each human life. This experimentation is done to verify and extend knowledge about a system. However, still many people are not aware that answers to problems arising from the way a system works may be obtained through experimentation, i.e. conducting experiments on the system by means of an experimentation system composed of measuring equipment. This way of experimentation covers all human and technical activities to perform an experiment, that is, stimulation of a system through stimuli (inputs), measuring the response (outputs), interpretation of the relations between outputs and inputs, and decision-making with respect to the continuation of the experimentation. In system studies it is usual to employ system modelling in order to represent a system by a model that is better manageable because the relations between inputs, state and outputs are unambiguously described through formulas/equations. In physics, economics, etc., system modelling leads to physical models, economic models, etc. Systems from different application areas e.g. physics and chemistry may still behave theirselves in a similar way. Then detaching the model descriptions from the nature of their original application areas will result in the same type of mathematical model. The reason for the structural similarity in space and time of mathematical models in different disciplines is due to the similarity of laws applied to derive the system equations. The use of mathematical models in system modelling makes knowledge gained for a mathematical model in one area of application easily portable to other areas. Experimentation has gone on for centuries to gather knowledge about physical phenomena on systems of practical importance. To obtain valid results by experimentation, it always was necessary to pay careful attention to tools and techniques, measuring and observation methods, and ways of exploiting experimental results. Precise and accurate experimentation has become a tradition. Consequently everyone considers results of experimentation on systems to be credible. In many cases it is economically not feasible or not possible to experiment on a system itself. Occasionally in the past much has been learned by substituting an other system for the actual

Optical interconnects in high-bandwidth computing

High bandwidth computing makes equally high demands upon the speed of data processing and the speed of data flow. Consequently, in high bandwidth computing systems at several levels, fast and highly parallel interconnects are necessary for both clock and data distribution in order to avoid transfer-bound computing. For the present time, data processing has still to be based upon the use of electronic components. However, realization of fast and highly parallel interconnects is no longer possible without application of optical (and electrical) interconnects in combination with electro-optic and opto-electronic elements. Realization problems are quite different at system, inter-chip and intra-chip level. High bandwidth computing requires at system level application of powerful, reconfigurable interconnects. Consequently, the feasibility of optical interconnects at the system level is closely related to the system interconnect topology. The advantages of the use of optical interconnects for the purpose of data distribution has given rise to consideration of the feasibility of replacing electrical power distribution by means of electrical links with optical power distribution through optical links.