Richard Miller and Joy Reed. The Oxford BSP Library Users' Guide, version 1.0. Technical report, Oxford Parallel, 1993.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....etc. Their analysis suggests that knowledge about the underlying struc ture of the problem is important to achieving efficient parallel computations on a 8 BSP computer. Several people are working on providing suitable environments for BSP pro gramming. Miller developed the Oxford BSP library [48, 49], a simple yet robust library for BSP programming. The library uses an SPMD model with static al location of processors. BSPlib [25] is another programming library based on the BSP model developed by the Oxford Parallel group. BSPlib is a part of the Ox ford BSP toolset which includes profiling ....

....until all of the data arrives at a processor. This feature may make logarithmic gather more attractive than linear gather under some circumstances. For example, using logarithmic gather, summation can be performed on the data being gathered. 3.1. 2 The Oxford BSP Library The Oxford BSP Library [48, 49], developed by Richard Miller, is used to implement the plasma simulation on NOWs. It is based on a slightly simplified version of the model presented in [70] The programming model uses a SPMD program style, and static allocation of processors. The most significant feature of the library is its ....

[Article contains additional citation context not shown here]

Richard Miller and Joy Reed. The Oxford BSP Library Users' Guide, version 1.0. Technical report, Oxford Parallel, 1993.

....has a SPARC processor with a clock speed of 33 Mhz. The computers are connected via Ethernet with a maximal throughput of 10 Mbit s. The computations have been performed in single precision arithmetic ( 8 significant digits) The communication has been implemented using the Oxford BSP library [3], according to the BSP programming model [1] 10] We selected the iterative solver BiCGstab(4) 6] with special techniques to improve the convergence properties [8] and the accuracy of the solution [7] For the preconditioner we have used (9) on the irregular grid, and (10) for the regular ....

R. Miller and J. Reed. The Oxford BSP library users' guide, version 1.00x:for'd Par'allel, Oxford 1993

....from free to unavailable as a transient failure. Another assumption of our approach is that the frequency of synchronization of parallel computation is high compared with the average available nonavailable times of a workstation. In this paper, we propose extensions to the Oxford BSP Library [6, 7] that allow for transparent recovery from transient processor failures in a network of workstations using dynamic replication of computation state, lazy process replication and process migration. The resulting system supports parallel computation with the level of parallelism changing in response ....

....main cost of which is the communication needed to send the state of each subcomputation to peer processes. 2. 2 Extensions to the Oxford BSP Library The above described scheme of using lay replication and migration to support adaptive parallelism has been integrated into the Oxford BSP library [6, 7]. It implements a simplified version of the Bulk Synchronous Parallel model [12] introduced by Leslie Valiant. It is simple, yet robust and was successfully used by us for implementing plasma simulation on a network of workstations. The extensions to the Oxford BSP library are introduced in two ....

Richard Miller and Joy Reed. The Oxford BSP Library Users' Guide, version 1.0. Technical report, Oxford Parallel, 1993.

....parameterises system characteristics in order to realise performance across distributed platforms without tying application code to a specific architecture. Predicting performance and determining optimal parallel strategies for applications are based on a corresponding cost model. The library [9] consists of a small number of subroutines to implement process creation, remote data access, and bulk synchronisation. The library can be linked to programs written in standard sequential languages such as Fortran, C, and Pascal. PORTABLE INTERFACE SINGLE SOURCE APPLICATION CODE The library ....

R Miller and J Reed, The Oxford BSP Library: Users' Guide, version 1.0, Oxford Parallel, 1993.

....Research Programme of the EC under contract N. 9072 (GEPPCOM project) and by CICYT project TIC96 0718 C02 01. 1 The library embodies a static SPMD (Single Program, Multiple Data) programming model, in which every parallel process executes the same program text; further details are given in [10]. To compute the predicted BSP cost on a given platform we determined the value of the BSP parameters for the Convex SPP1000 and the IBM SP2 using the notions introduced in [5] according to Miller s refinement of the BSP cost model [11] The paper is organized as follows. In Section 2 we recall ....

R. Miller and J. L. Reed. The Oxford BSP library users' guide. Technical report, Programming Research Group, University of Oxford, 1993.

....in Table 2 we considered the worst case behaviour, namely, swapping two columns is needed at each stage of the factorization process. 4 Comparision with the Experimental Results Obtained on an IBM SP2 We implemented the algorithms described in the previous section using the Oxford BSP library [13, 14]. This library callable from C, of FORTRAN, consists of a small set of subroutines that allow process management, remote data access and bulk synchronisation. The library has been implemented for different platforms ranging from networks of workstations to shared memory multiprocessors and ....

R. Miller and J. L. Reed. The Oxford BSP library users' guide. Technical report, Programming Research Group, University of Oxford, 1993.

....[13] 8] 2] to predict the performance of parallel algorithms. With this purpose we analyse the well known Preconditioned Conjugate Gradient (PCG) Method, with three different parallel preconditioners.We implemented these algorithms on a IBM SP2 with 8 processors using Oxford BSP library [9] [10]. This paper is organized as follows. In section 2 we recall the main aspects of the BSP model. In section 3 some fundamental ideas of the PCG method are explained briefly. We also analyse the theoretical BSP cost of the method for the different preconditioners used in section 4. Then, in section ....

R. Miller and J. L. Reed. The Oxford BSP library users' guide. Technical report, Programming Research Group, University of Oxford, 1993.

....typically requires a larger input size for BSP code than for machine specific code. A formal analysis of some BSP algorithms can be found in [GV92] A rather nice example of how a programmer can use the parameters p, L, g, and s, to optimize code is parameterizing for performance, given in [MR94] Depending on the values of the performance parameters, it may be less costly to broadcast data to processors linearly than to broadcast logarithmically. With values of p, L, g, and s for a specific machine available at compile time, preprocessor directives could cause compilation of that code ....

Richard Miller and Joy Reed. The Oxford BSP Library Users' Guide. Oxford Parallel, Oxford University Computing Laboratory, Wolfson Building, Parks Road, Oxford, OX1 3QD, 1.0 edition, May 1994.

....from remotely communicated data . This ensures that the semantics of a superstep becomes independent of when remote communications are completed within a superstep. Thus, the definition by McColl is included as one possibility. Further details can be found in The Oxford BSP Library User s Guide [MR95]. The current version of BSPlab performs all remote memory accesses at the end of a superstep. Later versions will allow experimenting with various implementation alternatives as mentioned above. Components and synchronization Both Valiant and McColl write that the periodic synchronization ....

....operation on values stored in registers or the local memory of that processor. For a given kind of processor, the MIPS rate should give a good estimate for the length of a time step and thus also the processor speed. For scientific applications it is common to take one flop as a (basic) time step [MR95]. Global operations are those that involve access to data not stored local to a processor. The time used to process such remote accesses should (as seen by the programmer) be independent of its location, i.e. a BSP computer has a twolevel memory model. This choice is crucial for making BSP ....

[Article contains additional citation context not shown here]

Richard Miller, and Joy Reed, The Oxford BSP Library Users' Guide, Oxford Parallel, Oxford University Computing Laboratory, 1.0 edition, May 1994.

....Parallel (BSP) 37] model of parallel computing offers an easy entry to parallel computing, and does not require great skill in parallel computing. It is a SIMD (Single Instruction Multiple Data) parallel model which is easily implemented through a library such as that developed by Miller and Reed [30]. The library contains only six operations as shown in Table 5. BSP START start of the BSP program BSP FINISH end of the BSP program BSP SSTEP START(n) start of superstep n BSP SSTEP END(n) end of superstep n BSP STORE(to, from data, to data, length) store from local to remote processor BSP ....

Richard Miller and Joy Reed. The Oxford BSP Library users' guide. Technical report, University of Oxford, 1993.

....of a 150MHz Dec Alpha processor and a memory of 64 Mbyte. Both communication networks of CrayT3D and Parsytec GC PowerPlus are threedimensional torus. All experimental machines are transformed into BSP architectures by using the corresponding implementations of the Oxford BSP library BSPLIB [16]. Usually the different sequential computing speeds s for different machines are obtained by measuring the time of a vector update operation. Values of the BSP cost model parameters with 32 bits reals as data word are shown in Table I. From the table, not very surprisingly, the two values of g, ....

R. Miller and J. Reed. The Oxford BSP library users' guide. Oxford Parallel, Oxford University, 1st edition, 1993.

....of fully portable (i.e. single source) parallel software which will run optimally on distributed memory architectures, shared memory multiprocessors, and on networks of workstations. A number of BSP software tools are being developed to support this work. One of these is the Oxford BSP Library [8]. The Library embodies a static SPMD (single program, multiple data) programming model, in which every parallel process executes the same program text. Processes proceed together through a sequence of supersteps; although within a superstep different processes may take different execution paths, ....

R Miller and J L Reed, The Oxford BSP Library: Users' Guide. Version 1.0, Oxford Parallel Technical Report, Oxford University Computing Laboratory, 1994.

....the algorithm on that computer. Of course, the accuracy of the prediction depends on how the BSP cost function reflects reality, and this may differ from machine to machine. Efficient implementations of the BSP model are currently being developed. One such implementation is the Oxford BSP library [5]. This public domain library is available for many architectures, including clusters of UNIX workstations, shared memory multiprocessors such as the Silicon Graphics Challenge, and massively parallel computers with distributed memory such as the Cray T3D. To give an impression of the wide range of ....

Miller, R., Reed, J.: The Oxford BSP library users' guide, version 1.0; Oxford Parallel, Oxford 1993.

....used for this paper are shown in Figure 6. 6 Speedup Prediction Previous reference was made in section 3 to the BSP parameters, and the fact that they could be used to make approximate predictions of the speedup which could be achieved by parallelising the GP process. In the Oxford BSP user guide[5], it is suggested that a conservative upper bound for the overhead involved in performing one of the supersteps is: 1 2 3 4 5 6 8 7 0 0 1 2 3 4 5 6 7 (i) Ring Topology (ii) Star Topology Fig. 6. Topologies used with the island model GP implementation. C L g Theta h : 1) The variables L and ....

Richard Miller and Joy Reed. The Oxford BSP Library users' guide. Technical report, University of Oxford, 1993.

....of the performance of the adaptive replication scheme along with the analysis can be found in (Nibhanupudi, 1998) 1.5 DESIGN AND IMPLEMENTATION OF A BSP LIBRARY 1.5. 1 Design of Adaptive Bulk Synchronous Parallel Library The adaptive replication scheme is developed using the Oxford BSP library (Miller, 1993; Miller and Reed, 1993) The library has been extended to pro 8 vide dynamic process management and virtual synchronization as described in (Nibhanupudi and Szymanski, 1996) Using the extended library, processes can now be terminated at any time or migrated to new host machines; new processes ....

....of the adaptive replication scheme along with the analysis can be found in (Nibhanupudi, 1998) 1.5 DESIGN AND IMPLEMENTATION OF A BSP LIBRARY 1.5. 1 Design of Adaptive Bulk Synchronous Parallel Library The adaptive replication scheme is developed using the Oxford BSP library (Miller, 1993; Miller and Reed, 1993). The library has been extended to pro 8 vide dynamic process management and virtual synchronization as described in (Nibhanupudi and Szymanski, 1996) Using the extended library, processes can now be terminated at any time or migrated to new host machines; new processes can be created to join ....

[Article contains additional citation context not shown here]

Miller, R. and Reed, J. (1993). The Oxford BSP Library Users' Guide, version 1.0. Technical report, Oxford Parallel.

....from free to unavailable as a transient failure. Another assumption of our approach is that the frequency of synchronization of parallel computation is high compared with the average available nonavailable times of a workstation. In this paper, we propose extensions to the Oxford BSP Library [6, 7] that allow for transparent recovery from transient processor failures in a network of workstations using dynamic replication of computation state, lazy process replication and process migration. The resulting system supports parallel computation with the level of parallelism changing in response ....

....main cost of which is the communication needed to send the state of each subcomputation to peer processes. 2. 2 Extensions to the Oxford BSP Library The above described scheme of using lazy replication and migration to support adaptive parallelism has been integrated into the Oxford BSP library [6, 7]. It implements a simplified version of the Bulk Synchronous Parallel model [12] introduced by Leslie Valiant. It is simple, yet robust and was successfully used by us for implementing plasma simulation on a network of workstations. The extensions to the Oxford BSP library are introduced in two ....

Richard Miller and Joy Reed. The Oxford BSP Library Users' Guide, version 1.0. Technical report, Oxford Parallel, 1993.

....direct mode the programmer retains control over data distribution and manipulation among the processors. A number of researchers are currently forming a World Wide Standard BSP Library [61] by synthesising several low level BSP programming approaches that have been pursued over the last few years [62, 99 101, 107]. They propose a library called BSPlib to provide a parallel communication library based around a SPMD model of computation. The main parts of BSPlib are: 1) routines to spawn a number of processes; 2) an operation that synchronises all processes; 3) Direct Remote Memory Access (DRMA) ....

R. Miller and J. Reed. The Oxford BSP Library Users' Guide, version 1.0. Technical report, Oxford Parallel, 1993.

....by the sentence While a PRAM language would be ideal, other styles also may be appropriate. in [Val90 p. 104] Languages based on the PRAM and BSP models are being further elaborated in the search for easy and efficient parallel programming [Rav95] Implementing the Oxford BSP library [Mil94, MR95] on a simulator will make it possible to execute real BSP programs 2 . Real applications are crucial for generating more realistic traces of computation and communication from the simulator than that can be achieved from toy examples and synthetic loads. 1 The parallel programming was done in a ....

....from remotely communicated data . This ensures that the semantics of a superstep becomes independent of when remote communications are completed within a superstep. Thus, the definition by McColl is included as one possibility. Further details can be found in The Oxford BSP Library User s Guide [MR95]. In the simulator, the first version will perform all remote memory accesses at the end of a superstep. Later versions will allow experimenting with various implementation alternatives as mentioned above. 2.3.3 Other differences Synchronize all components. Both Valiant and McColl write that the ....

[Article contains additional citation context not shown here]

Richard Miller, and Joy Reed, The Oxford BSP Library Users' Guide, Oxford Parallel, Oxford University Computing Laboratory, 1.0 edition, May 1994.

....We briefly introduce the current direct mode BSP programming environment before discussing our approach. 6. 1 BSPlib: The World Wide BSP Programming Environment By synthesising several low level BSP programming approaches that have been pursued over the last few years [GRT94, RPL95, Mil93, MR93, Mil94] a World Wide Standard BSP Library [GHL 96] has been formulated. The proposed library, called BSPlib, provides a parallel communication facilities based around a SPMD model of computation. The main parts of BSPlib are: 1) routines to spawn a number of processes; 2) an operation that ....

R. Miller and J. Reed. The Oxford BSP Library Users' Guide, version 1.0. Technical report, Oxford Parallel, 1993.

....of fully portable (i.e. single source) parallel software which will run optimally on distributed memory architectures, shared memory multiprocessors, and on networks of workstations. A number of BSP software tools are being developed to support this work. One of these is the Oxford BSP Library [13]. The Library embodies a static SPMD (single program, multiple data) programming model, in which every parallel process executes the same program text. Processes proceed together through a sequence of supersteps; although within a superstep different processes may take different execution paths, ....

R Miller and J L Reed. The Oxford BSP Library: Users' Guide. Version 1.0. Oxford Parallel Technical Report, Oxford University Computing Laboratory, 1994.

....systems is under development. 10 ALTERNATIVE PROGRAMMING MODELS 44 Contact: Joy Reed, Oxford Parallel, OUCL Wolfson House, Parks Road, Oxford, OX1 3QD, UK. Email: joy comlab.ox.ac.uk FTP: Comments: Information available on WWW http: www.comlab.ox. ac.uk oucl oxpara bsplib1.html References: [96, 97] 10.11 P4 LINDA Name: P4 LINDA Description: Package of C routines which provide some of the core functions of the Linda programming model as described in [64] Systems: Sits on top of the P4 communications harness (see separate entry) Can make use of the message passing facilities of P4 on a ....

R Miller and J Reed. The Oxford BSP Library: Users Guide. Oxford Parallel, version 1.0 edition, 1993.

No context found.

Richard Miller and Joy Reed. The oxford bsp library users' guide, version 1.0. Technical report, Oxford Parallel, 1993.

No context found.

R. Miller and J. Reed, The Oxford BSP Library: Users Guide, Version 1.0, Oxford Parallel (1993).

No context found.

Richard Miller and Joy Reed. The Oxford BSP Library users' guide. Technical report, University of Oxford, 1993.

No context found.

R. Miller and J. Reed, The Oxford BSP Library: Users Guide, Version 1.0, Oxford Parallel (1993).

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC