Validity of The Single Processor Approach to Achieving Large Scale Computing Capabilities | CSC 8400, Papers of Operating Systems

Download Validity of The Single Processor Approach to Achieving Large Scale Computing Capabilities | CSC 8400 and more Papers Operating Systems in PDF only on Docsity! AFIPS spring joint computer conference Validity of the single processor approach to achieving large scale computing capabilities Gene M Amdahl IBM Sunnyvale California  INTRODUCTION For over a decade prophets have voiced the contention that the organization of a single computer has reached its limits and that truly signicant advences can be made only by interconnection of a multiplicity of computers in such a manner as to permit cooperative solution Variously the proper direction has been pointed out as general purpose computers with a generalized interconnection of memories or as specialized computers with geometrically related memory interconnections and controlled by one or more instruction streams Demonstration is made of the continued validity of the single processor approach and of the weaknesses of the multiple processor approach in terms of application to real problems and their attendant irregularities The arguments presented are based on statistical characteristics of computation on computers over the last decade and upon the operational requirements within problems of physical interest An additional reference will be one of the most thorough analyses of relative computer capabilities currently published Changes in Computer Performance Datamation September   Professor Kenneth F Knight Stanford School of Business Asministration The rst characteristic of interest is the fraction of the computational load which is associated with data management housekeeping This fraction has been very nearly constant for about ten years and accounts for of the executed instructions in production runs In an entirely dedicated special purpose environment this might be reduced by a factor of two but it is highly improbably that it could be reduced by a factor of three The nature of this overhead appears to be sequential so that it is unlikely to be amenable to parallel processing techniques Overhead alone would then place an upper limit on throughput of ve to seven times the sequential processing rate even if the This paper is retyped as the present form by Guihai Chen He wishes you would enjoy reading this historical paper  housekeeping were done in a separate processor The non housekeeping part of the problem could exploit at most a processor of performance three to four times the performance of the housekeeping processor A fairly obvious conclusion which can be drawn at this point is that the e ort expended on achieving high parallel processing rates is wasted unless it is accompanied by achievements in sequential processing rates of very nearly the same magnitude Data management housekeeping is not the only problem to plague oversimplied approaches to high speed computation The physical problems which are of practical interest tend to have rather signicant complications Examples of these complications are as follows boundaries are likely to be irregular interiors are inhomogeneous computations required may be dependent on the states of the variables at each point propagation rates of di erent physical e ects may be quite di erent the rate of convergence or convergence at all may be strongly dependent on sweeping through the array along di erent axes on succeeding passes etc The e ect of each of these complications is very severe on any computer organization based on geometrically related processors in a paralleled processing system Even the existence of regular rectangular boundaries has the interesting property that for spatial dimension of N there are N di erent point geometries to be dealt with in a nearest neighbor computation If the second nearest neighbor were also involved there would be N di erent point geometries to contend with An irregular boundary compounds this problem as does an inhomogeneous interiors Computations which are dependent on the states of variables would require the processing at each point to consume approximately the same computational times as the sum of computations of all physical e ects within a large region Di erences of changes in propagation rates may a ect the mesh point relationships Ideally the computation of the action of the neighboring points upon the point under consid eration involves their values at a previous time proportional to the mesh spacing and inversely proportional to the propagation rate Since the time step is normally kept constant a faster prop agation rate for some e ects would imply interactions with more distant points Finally the fairly common practice of sweeping through the mesh along di erent axes on succeeding passes posed problems of data management which a ects all processors however it a ects geometrically related processors more severely by requiring transposing all points in storage in addition to the revised inputoutput scheduling A realistic assessment of the e ect of these irregularities on a simplied and regularized abstraction of the problem yields a degradation in the vicinity of onehalf to one order of magnitude To sum up the e ects of data management housekeeping and of problem irregularities the author has compared three di erent machine organizations involving approximately equal amounts