Pub Date : 1991-04-28DOI: 10.1109/DMCC.1991.633349
L. S. Kaplan
High bandwidth delivery of data to the processor(s) is critical for good perforniance in highly parallel computer systems. To increase memory throughput, many systems make use of interleaved parallel memory banks. An implementation must provide uniform throughput with little or no contention at the memory banks for a wide variety of algorithms and access patterns. This paper proposes an implementation for an interleaved memory system that exhibits extremely low contention for the memoiry banks during virtually all patterned accesses. It also has the advantage that, due to its programmability, it imposes few requirements on the configuration of the machines in which it is used. The hardware to implement the design is dliscussed along with address space considerations. A variant of this design is currently in use on the BBN TC2000 (tm) parallel computer.
{"title":"A Flexible Interleaved Memory Design for Generalized Low Conflict Memory Access","authors":"L. S. Kaplan","doi":"10.1109/DMCC.1991.633349","DOIUrl":"https://doi.org/10.1109/DMCC.1991.633349","url":null,"abstract":"High bandwidth delivery of data to the processor(s) is critical for good perforniance in highly parallel computer systems. To increase memory throughput, many systems make use of interleaved parallel memory banks. An implementation must provide uniform throughput with little or no contention at the memory banks for a wide variety of algorithms and access patterns. This paper proposes an implementation for an interleaved memory system that exhibits extremely low contention for the memoiry banks during virtually all patterned accesses. It also has the advantage that, due to its programmability, it imposes few requirements on the configuration of the machines in which it is used. The hardware to implement the design is dliscussed along with address space considerations. A variant of this design is currently in use on the BBN TC2000 (tm) parallel computer.","PeriodicalId":313314,"journal":{"name":"The Sixth Distributed Memory Computing Conference, 1991. Proceedings","volume":"22 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1991-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"116122453","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 1991-04-28DOI: 10.1109/DMCC.1991.633216
K. Dalton, E. Hensel, S. Castillo, K. Ng
A study of the finite difSerence solution of the nonlinear partial differential equations governing twoand three-dimensional semiconductor devices is conducted on a SIMD computer. This nonlinear system is solved using Jacobi iteration and successive-under-relaxation. Row scaling and a zero order regularizer are used to aid in convergence. On a 16K CM-2 problems with up to 16.7 million unknowns have been solved. Problems of this size have not previously been reported. The ability to accurately model larger and more realistic three-dimensional devices is necessary to gain a greater physical understanding of their behavior.
{"title":"The Finite Difference Solution of Two- and Three-Dimensional Semiconductor Problems on the Connection Machine","authors":"K. Dalton, E. Hensel, S. Castillo, K. Ng","doi":"10.1109/DMCC.1991.633216","DOIUrl":"https://doi.org/10.1109/DMCC.1991.633216","url":null,"abstract":"A study of the finite difSerence solution of the nonlinear partial differential equations governing twoand three-dimensional semiconductor devices is conducted on a SIMD computer. This nonlinear system is solved using Jacobi iteration and successive-under-relaxation. Row scaling and a zero order regularizer are used to aid in convergence. On a 16K CM-2 problems with up to 16.7 million unknowns have been solved. Problems of this size have not previously been reported. The ability to accurately model larger and more realistic three-dimensional devices is necessary to gain a greater physical understanding of their behavior.","PeriodicalId":313314,"journal":{"name":"The Sixth Distributed Memory Computing Conference, 1991. Proceedings","volume":"127 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1991-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"127401838","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 1991-04-28DOI: 10.1109/DMCC.1991.633209
R. Firestone, Eric N. Opp
The performance of reflecting optical telescopes located on the surface of the earth are subject to distortions due to the force of gravity on the mirror and the turbulence of the atmosphere on the light path. Reflective optics are also planned for use in high-powered laser systems, where the intensity of the light itself is capable of producing distortions in the air within the instrument, thereby affecting the shape of the focused wavefront. A solution proposed by optical designers is the use of adaptive optics: an optical system in which the figure of the mirror is deformable to the extent necessary to correct for the distortions mentioned. An adaptive optical system uses a feedback loop concept, in which the distortions of the optical wavefront are measured, the necessary corrections are computed, and a set of actuators is moved to provide those corrections. The calculation of the corrections is computationally intense. Specifically, the measurement of the distortions provides a collection of phase differences between measuring points corresponding to the actuator positions. This set of phase differences is larger than the number of actuators, leading to an overdetermined problem. As physical systems have some amount of noise present, the technique of least-squares solution serves both to provide the best choice of actuator positions for this overdetermined problem and to suppress the noise in the measurements. The necessary algorithms for solving the computation portion of the adaptive optics problem consist of a matrix generator to derive the computational representation of the physical system, a matrix inversion routine, and a high-speed least-squares solver. In the optical astronomy paradigm, the computational requirement is for a small number of adjustments per second, due to the rate of atmospheric turbulence. For the laser system, with more stringent requirements, we demonstrate an improvement of 11 2 orders of magnitude, made possible only through the use of supercomputer methods. Extrapolation of these results indicates that even greater acceleration is possible if the interprocessor communication is minimized; in other words, supercomputer designers have not yet solved the problem of making interprocessor communication as efficient as that within processors (or, in the present case, between processors on a single chip).
{"title":"Adaptive Optics Calculations Using the Connection Machine","authors":"R. Firestone, Eric N. Opp","doi":"10.1109/DMCC.1991.633209","DOIUrl":"https://doi.org/10.1109/DMCC.1991.633209","url":null,"abstract":"The performance of reflecting optical telescopes located on the surface of the earth are subject to distortions due to the force of gravity on the mirror and the turbulence of the atmosphere on the light path. Reflective optics are also planned for use in high-powered laser systems, where the intensity of the light itself is capable of producing distortions in the air within the instrument, thereby affecting the shape of the focused wavefront. A solution proposed by optical designers is the use of adaptive optics: an optical system in which the figure of the mirror is deformable to the extent necessary to correct for the distortions mentioned. An adaptive optical system uses a feedback loop concept, in which the distortions of the optical wavefront are measured, the necessary corrections are computed, and a set of actuators is moved to provide those corrections. The calculation of the corrections is computationally intense. Specifically, the measurement of the distortions provides a collection of phase differences between measuring points corresponding to the actuator positions. This set of phase differences is larger than the number of actuators, leading to an overdetermined problem. As physical systems have some amount of noise present, the technique of least-squares solution serves both to provide the best choice of actuator positions for this overdetermined problem and to suppress the noise in the measurements. The necessary algorithms for solving the computation portion of the adaptive optics problem consist of a matrix generator to derive the computational representation of the physical system, a matrix inversion routine, and a high-speed least-squares solver. In the optical astronomy paradigm, the computational requirement is for a small number of adjustments per second, due to the rate of atmospheric turbulence. For the laser system, with more stringent requirements, we demonstrate an improvement of 11 2 orders of magnitude, made possible only through the use of supercomputer methods. Extrapolation of these results indicates that even greater acceleration is possible if the interprocessor communication is minimized; in other words, supercomputer designers have not yet solved the problem of making interprocessor communication as efficient as that within processors (or, in the present case, between processors on a single chip).","PeriodicalId":313314,"journal":{"name":"The Sixth Distributed Memory Computing Conference, 1991. Proceedings","volume":"229 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1991-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130910711","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 1991-04-28DOI: 10.1109/DMCC.1991.633198
J. Mcdonald, L. Dagum
Direct particle simur'ation is a powerful method for analyzing low density, hypersonic re-entry flows. The method involves following a large sample of representative gas molecules through motion and collision with other molecules or with surfaces in the simulated flow. In this paper, two very different parallel architectures are examined for their suitability an particle samulation computations, na;mely the Connection Machine CM-2 and the Intel iPSC/860. The difference in architectures has resulted in very diferent parallel decompositions. The two implementations are described and performance results are given. Both implementations achieve performance comparable iio a single Cray-2 CPU, however, this performance is obtained at the cost of greatly increased programming complexity.
{"title":"A Comparison of Particle Simulation Implementations on Two Different Parallel Architect ures","authors":"J. Mcdonald, L. Dagum","doi":"10.1109/DMCC.1991.633198","DOIUrl":"https://doi.org/10.1109/DMCC.1991.633198","url":null,"abstract":"Direct particle simur'ation is a powerful method for analyzing low density, hypersonic re-entry flows. The method involves following a large sample of representative gas molecules through motion and collision with other molecules or with surfaces in the simulated flow. In this paper, two very different parallel architectures are examined for their suitability an particle samulation computations, na;mely the Connection Machine CM-2 and the Intel iPSC/860. The difference in architectures has resulted in very diferent parallel decompositions. The two implementations are described and performance results are given. Both implementations achieve performance comparable iio a single Cray-2 CPU, however, this performance is obtained at the cost of greatly increased programming complexity.","PeriodicalId":313314,"journal":{"name":"The Sixth Distributed Memory Computing Conference, 1991. Proceedings","volume":"69 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1991-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"131751130","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 1991-04-28DOI: 10.1109/DMCC.1991.633097
J. Yantchev
Concurrent systems are collections of data, processes, and communication channels. Top-down, hierarchical design of concurrent systems needs powerful abstraction facilities provided by the implementation language. While most languages provide some structuring mechanisms for data and process abstraction, none seems to provide any equivalent mechanisms for communication structuring. Communication channels are to communicate data and, therefore, all data structuring mechanisms provided by a programming language must be available to structure channels as well. In order to preserve behaviour through successive levels of design refinement, these means of communication structuring must preserve the abstraction of atomic transfers of values of arbitrary types. Int r o duct ion Most concurrent programming languages [5, 6, 7, 11 support the abstraction of concurrent systems as collections of data, processes, and communication channels. However, while they provide some structuring mechanisms for data and process abstraction, none seems to provide any equivalent mechanisms for communication structuring. Interprocess communication is almost universally viewed as a synchronised atomic exchange of values between two concurrently active processes. This affects the whole design process and intervenes with the freedom and ease in the refinement of the process structure. The design transformation steps may be non-trivial in some cases and, therefore, difficult to arrive at and verify. In addition, the implementation may be less efficient, both in storage and speed, because of unnecessary data copying and context creation for process spawning. The data structuring mechanisms supported by the contemporary programming languages provide a uniform view on data and data types. Structured data types may consist of components of arbitrary types, including themselves, and values of such types are treated as wholes and may be passed as parameters, returned as results of functions, and assigned to variables. The same applies to processes [5, 71. No distinction of kind need be made between systems with and without substructure and, indeed, a system which at one level of abstraction may be considered to consist of a process and the environment in which it evolves, may be considered as a single system at a higher level of abstraction. A process which for one purpose is taken to be atomic
{"title":"Communication Abstraction and Process Refinement","authors":"J. Yantchev","doi":"10.1109/DMCC.1991.633097","DOIUrl":"https://doi.org/10.1109/DMCC.1991.633097","url":null,"abstract":"Concurrent systems are collections of data, processes, and communication channels. Top-down, hierarchical design of concurrent systems needs powerful abstraction facilities provided by the implementation language. While most languages provide some structuring mechanisms for data and process abstraction, none seems to provide any equivalent mechanisms for communication structuring. Communication channels are to communicate data and, therefore, all data structuring mechanisms provided by a programming language must be available to structure channels as well. In order to preserve behaviour through successive levels of design refinement, these means of communication structuring must preserve the abstraction of atomic transfers of values of arbitrary types. Int r o duct ion Most concurrent programming languages [5, 6, 7, 11 support the abstraction of concurrent systems as collections of data, processes, and communication channels. However, while they provide some structuring mechanisms for data and process abstraction, none seems to provide any equivalent mechanisms for communication structuring. Interprocess communication is almost universally viewed as a synchronised atomic exchange of values between two concurrently active processes. This affects the whole design process and intervenes with the freedom and ease in the refinement of the process structure. The design transformation steps may be non-trivial in some cases and, therefore, difficult to arrive at and verify. In addition, the implementation may be less efficient, both in storage and speed, because of unnecessary data copying and context creation for process spawning. The data structuring mechanisms supported by the contemporary programming languages provide a uniform view on data and data types. Structured data types may consist of components of arbitrary types, including themselves, and values of such types are treated as wholes and may be passed as parameters, returned as results of functions, and assigned to variables. The same applies to processes [5, 71. No distinction of kind need be made between systems with and without substructure and, indeed, a system which at one level of abstraction may be considered to consist of a process and the environment in which it evolves, may be considered as a single system at a higher level of abstraction. A process which for one purpose is taken to be atomic","PeriodicalId":313314,"journal":{"name":"The Sixth Distributed Memory Computing Conference, 1991. Proceedings","volume":"102 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1991-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"133378155","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 1991-04-28DOI: 10.1109/DMCC.1991.633360
B. Izadi, F. Ozguner
This paper investigates hardware reconjiguratzon schemes to make the hypercube multicomputer fault tolerant. Two schemes are proposed; the Cluster Approach and the Enhanced Cluster Approach. The approaches are shown to be able to tolerate large number of failures without any performance deg,radation. It is further demonstrated that no modification to either the existing communication or computaitional algorithm is needed. Finally a gracefully degmdable approach is presented to reconfigure when the number of faulty nodes are more than the available spares.
{"title":"Spare Allocation and Reconfiguration in a Fault Tolerant Hypercube with Direct Connect Capability","authors":"B. Izadi, F. Ozguner","doi":"10.1109/DMCC.1991.633360","DOIUrl":"https://doi.org/10.1109/DMCC.1991.633360","url":null,"abstract":"This paper investigates hardware reconjiguratzon schemes to make the hypercube multicomputer fault tolerant. Two schemes are proposed; the Cluster Approach and the Enhanced Cluster Approach. The approaches are shown to be able to tolerate large number of failures without any performance deg,radation. It is further demonstrated that no modification to either the existing communication or computaitional algorithm is needed. Finally a gracefully degmdable approach is presented to reconfigure when the number of faulty nodes are more than the available spares.","PeriodicalId":313314,"journal":{"name":"The Sixth Distributed Memory Computing Conference, 1991. Proceedings","volume":"23 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1991-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"123649585","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 1991-04-28DOI: 10.1109/DMCC.1991.633106
Minyou Wu, W. Shu
An extended scatter scheduling was applied to problems with unpredictable, asynchronous struc- tures. It has been found that with this simple schedul- ing strategy, good load balance can be reached with- out incurring much runtime overhead. This scheduling algorithm has been implemented on hypercube ma- chines, and its performance is compared with other scheduling strategies.
{"title":"Scatter Scheduling for Problems with Unpredictable Structures","authors":"Minyou Wu, W. Shu","doi":"10.1109/DMCC.1991.633106","DOIUrl":"https://doi.org/10.1109/DMCC.1991.633106","url":null,"abstract":"An extended scatter scheduling was applied to problems with unpredictable, asynchronous struc- tures. It has been found that with this simple schedul- ing strategy, good load balance can be reached with- out incurring much runtime overhead. This scheduling algorithm has been implemented on hypercube ma- chines, and its performance is compared with other scheduling strategies.","PeriodicalId":313314,"journal":{"name":"The Sixth Distributed Memory Computing Conference, 1991. Proceedings","volume":"65 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1991-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"130363949","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 1991-04-28DOI: 10.1109/DMCC.1991.633075
M. Livingston, Q. Stout
This paper examines the problem of locating large fault-free subcubes in multiuser hypercube systems. We analyze a new location strategy, the cyclic buddy system, and compare its performance to the buddy system, the gray-coded buddy system, and several variants of them. We show that the cyclic buddy system gives a striking improvement in expected fault tolerance over the above schemes and, since it can easily be implemented in parallel with little overhead, it provides an attractive alternative to these schemes. We also investigate the behavior of these location systems in the folded, or projective, hypercube, and find that the cyclic buddy system, which adapts naturally to this enhancement, significantly outperforms the other schemes. A combination of analytic techniques and simulation is used to examine both worst case and expected case performance.
{"title":"Fault Tolerance of the Cyclic Buddy Subcube Location Scheme in Hypercubes","authors":"M. Livingston, Q. Stout","doi":"10.1109/DMCC.1991.633075","DOIUrl":"https://doi.org/10.1109/DMCC.1991.633075","url":null,"abstract":"This paper examines the problem of locating large fault-free subcubes in multiuser hypercube systems. We analyze a new location strategy, the cyclic buddy system, and compare its performance to the buddy system, the gray-coded buddy system, and several variants of them. We show that the cyclic buddy system gives a striking improvement in expected fault tolerance over the above schemes and, since it can easily be implemented in parallel with little overhead, it provides an attractive alternative to these schemes. We also investigate the behavior of these location systems in the folded, or projective, hypercube, and find that the cyclic buddy system, which adapts naturally to this enhancement, significantly outperforms the other schemes. A combination of analytic techniques and simulation is used to examine both worst case and expected case performance.","PeriodicalId":313314,"journal":{"name":"The Sixth Distributed Memory Computing Conference, 1991. Proceedings","volume":"3 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1991-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121643148","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 1991-04-28DOI: 10.1109/DMCC.1991.633138
Bernd Bruegge, Hiroshi Nishikawa, Peter Steenkiste
With the advances in high-speed networking, partitioning applications over a group of computer systems is becoming an attractive way of exploiting parallelism. Programming general multicomputers is however very challenging: nodes are typically heterogeneous and shared with other users, making the availability of computing cycles on the nodes and communication bandwidth on the network unpredictable, This environment often requires users to use a programming model based on dynamic load balancing. In this paper, we use an flow field generation application to look at the problems that come up in a network environment. We use BEE, a monitoring system that allows programmers to interactively monitor their application, to show the behavior of the program under different conditions.
{"title":"Computing over Networks: An Illustrated Example","authors":"Bernd Bruegge, Hiroshi Nishikawa, Peter Steenkiste","doi":"10.1109/DMCC.1991.633138","DOIUrl":"https://doi.org/10.1109/DMCC.1991.633138","url":null,"abstract":"With the advances in high-speed networking, partitioning applications over a group of computer systems is becoming an attractive way of exploiting parallelism. Programming general multicomputers is however very challenging: nodes are typically heterogeneous and shared with other users, making the availability of computing cycles on the nodes and communication bandwidth on the network unpredictable, This environment often requires users to use a programming model based on dynamic load balancing. In this paper, we use an flow field generation application to look at the problems that come up in a network environment. We use BEE, a monitoring system that allows programmers to interactively monitor their application, to show the behavior of the program under different conditions.","PeriodicalId":313314,"journal":{"name":"The Sixth Distributed Memory Computing Conference, 1991. Proceedings","volume":"710 1","pages":"0"},"PeriodicalIF":0.0,"publicationDate":"1991-04-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"121998608","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：