F. Andr'e, J. Pazat, and H. Thomas. Pandore: A system to manage data distribution. In Proceedings of the  Home/Search   Document Not in Database   Summary   Related Articles   Check

....do not directly specify alignments between arrays. Instead, they distribute each array individually and implicitly derive the alignment between different arrays based on their relative distributions. Systems such as Dino [RSW90] Id Nouveau [RP89] Mimdizer [SWW91] Oxygen [RA90] and Pandore [APT90] all provide data distribution specifications equivalent to some combination of BLOCK and CYCLIC. Dino also supports special stencil based data distributions with overlaps. The Aspar [IFKF90] compiler performs automatic data decomposition and communications generation. A micro stencil is derived ....

F. Andr'e, J. Pazat, and H. Thomas. Pandore: A system to manage data distribution. In Proceedings of the

....defined. Several compilation systems for exploiting fine grained parallelism have been and are being built, which include Al [Tse90] Aspar [IFKF90] C Dataparallel C [HQL 91, RS87] Crystal [LC91] Dino [RSW91] Id Nouveau [RP89] Mimdizer [SWW92] Oxygen [RA90] P 3 C [GAY91] Pandore [APT90] Parafrase 2 [GB92] Paragon [CCRS91] Spot [SS90, Soc90] Superb [ZBG88] and Vienna Fortran [BCZ92] While there is still much work to be done in this field in general, there has already been considerable success in the field of regular applications, and second generation parallelizing ....

F. Andr'e, J. Pazat, and H. Thomas. Pandore: A system to manage data distribution. In Proceedings of the 1990 ACM International Conference on Supercomputing, Amsterdam, The Netherlands, June 1990.

....processes mapping, etc. Since then, researches have undertaken the task of providing parallel extensions to sequential languages such as C or Fortran to hide the distributed nature of these architectures. Several prototypes have been built, including Fortran D [10] Vienna Fortran [3] Pandore [2]. Recently, a coalition of industrial and academics group founded the High Performance Fortran Forum to propose extensions to Fortran (HPF) All these approaches provide the user with a global address space Permission to copy without fee all or part of this material is granted provided that the ....

....execution model, a process is created on each processor at the beginning of the program execution and then executed. These processes remain alive until the end of the program. This execution model has been widely investigated in the design of programming environments that exploits data parallelism [2, 3, 10]. However, few work has been done to combine both control parallelism and SPMD code generation. This idea has been investigated within the EPEX project [9, 21] but the user programming model is more complicated than ours because users have to work with several types of code sections: uniprocessor ....

Francoise Andr'e, Jean-Louis Pazat, and Henry Thomas. Pandore : A System to Manage Data Distribution. In International Conference on Supercomputing, ACM, June 11-15 1990.

....mapping, etc. Since then, researches have undertaken this task by providing parallel extensions to sequential languages such as C or Fortran to hide the distributed nature of these architectures. Several prototypes have been carried out as for instance Fortran D[4] Vienna Fortran [2] Pandore [1]. Recently, a coalition of industrial and academics group founded a forum (High Performance Fortran Forum) to propose parallel extensions to Fortran (HPF) All these approaches provide the user with a global address space and the compilers are in charge of generating processes and communications. ....

Francoise Andr'e, Jean-Louis Pazat, and Henry Thomas. Pandore : A system to manage Data Distribution. In International Conference on Supercomputing, ACM, June 11-15 1990.

....It necessary because Fortran remainder is not positive for negative numbers. It was added to insure the semantics of the output code. 6 Related work Techniques to generate distributed code from sequential or parallel code using a uniform memory space have been extensively studied since 1988 [9, 28, 40, 31, 15, 22, 16, 6, 25, 2, 21, 27, 32, 23, 5, 19, 4, 8, 11, 39]. The most obvious, most general and safest technique is called run time resolution [9, 28, 31] Each instruction is guarded by a condition that is true for processors that must execute it. Each memory address is checked before it is referenced to decide whether the address is local and the ....

Francoise Andr' e, Jean-Louis Pazat, and Henry Thomas. Pandore: A System to Manage Data Distribution. Publication Interne 519, IRISA, February 1990.

....Code invariance (Portability) Ref. SUPERB extra annotation transparent with regular parallelism [6] Vienna Fortran extra annotation transparent with regular parallelism [7] HPF extra annotation transparent with regular parallelism [8] PANDORE extra annotation transparent with regular parallelism [9] Crystal extra annotation compiler dependent [10] MENTAT mixed in a single class modification may be required [11] pC (old) mixed in a single class modification may be required [12] EPEE Paladin mixed in a single class modification may be required [13] ScaLAPACK mixed in a single class ....

Francoise Andr'e, Jean-Louis Pazat, and Henry Thomas. Pandore: a system to manage data distribution. In ICS '90, September 1990.

....to assembly language comparing to imperative language. Several alternatives have been already investigated to program DMPCs using a high level programming model. Most of them aim at to provide a global address space implemented either by a compiler (Fortran D [17] Vienna Fortran [4] Pandore [1] and HPF [11] or by an operating system (Shiva [21] koan [19] myoan [8] asvm [27] max [15] or Treadmarks [22] This latter approach extend the virtual memory concept to parallel architectures and is known under various names such as DSM (Distributed Shared Memory) or SVM (Shared Virtual ....

F. Andr'e, J.-L. Pazat, and H. Thomas. PANDORE: A system to manage data distribution. In Int. Conf. on Supercomputing, pages 380--388, June 1990.

.... machines can adjust to a changing number of PEs by redistributing their data structures [422, 171, 396] In some cases, such redistribution is supported by the system (albeit with the intention of redistribution for different phases of the computation, not for changes in the available PEs) [23, 378]. As this involves considerable overhead, it is imperative that the frequency of reconfigurations be kept low in practice: otherwise the price of reconfiguration might even outweigh the benefits of changing the PE allocation [446, 171, 543] Alternatively, a model where reconfigurations are only ....

....communication requires knowledge about the interactions among the threads. The PARMACS message passing library and MPI allow such interactions to be expressed via the virtual process topology [95, 604] HPF, Dataparallel C, and Pandore only support a more restrictive multidimensional mesh topology [378, 257, 23]. Alternatively, the system can make an educated guess about communication patterns; for example, it is reasonable to assume that when one thread spawns another thread, the two will communicate [56] However, in many cases the system does not have enough information, and the mapping has to be done ....

F. Andr'e, J-L. Pazat, and H. Thomas, "Pandore: a system to manage data distribution". In Intl. Conf. Supercomputing, pp. 380--388, Jun 1990.

....communication requires knowledge about the interactions among the threads. The PARMACS message passing library and MPI allow such interactions to be expressed via the virtual process topology [51, 351] HPF, Dataparallel C, and Pandore only support a more restrictive multidimensional mesh topology [227, 157, 15]. Alternatively, the system can make an educated guess about 32 communication patterns; for example, it is reasonable to assume that when one thread spawns another thread, the two will communicate [34] 4.1.2 Load balancing The second major issue involved in the use of local queues is load ....

F. Andr'e, J-L. Pazat, and H. Thomas, "Pandore: a system to manage data distribution". In Intl. Conf. Supercomputing, pp. 380--388, Jun 1990.

.... approach, which consists in automatically creating communicating processes from sequential code plus data decomposition specifications. The most usual code generation is based on the Single Program Multiple Data principle [6] and on the standard notion of refresh of remote variables [1]. The correctness of the distribution is achieved when, given corresponding initial data, either the source sequential code and the parallel generated one produce corresponding results, or both of them diverge [3] ffl Automated distribution of reactive systems: in that case, increasing ....

.... assignment is only replicated on the process which owns the left hand side variable (owner compute rule [6] In short, communication code is generated in order to implement access to remote variables: a copy of each referenced variable is to be sent by its owner towards the process that needs it [1, 3, 7]. However our only concern is distribution of control. Data exchange is not under consideration. Action S a i j (resp. S a i j ) is a message send (resp. receive) from process i to process j with value a. For all i 2 I, Delta i = Phi S a j i ; S a i j Psi a2 Sigma;j2I is the ....

Francoise Andr'e, Jean-Louis Pazat, and Henry Thomas. Pandore: a system to manage data distribution. In ACM International Conference on Supercomputing, June 11-15 1990. 15

....] N=P ] where P is the number of processors. The memory overhead induced by this layout remains small. Proposals have also been made to allow for alignment while keeping memory use at a minimum [6] To complement the management of local partitions, several ad hoc techniques (temporary scalars [2], buffers [11] hash tables [7] extension of local partition [12] are used for handling received data. The additional memory space required varies according to the chosen technique but must remain acceptable. 2 Yves Mah eo, Jean Louis Pazat 2.2 Global to Local Index Conversion In order to ....

F. Andr'e, J.L. Pazat, and H. Thomas. Pandore a system to manage data distribution. In International Conference on Supercomputing, ACM, June 1990.

.... early performance feedback in the upper, compile time layer of the performance prediction hierarchy has been applied in many parallel programming environments, in which parallelization is performed either explicitly [12, 35, 26] or through restructuring [2, 11, 20, 36] and data partitioning [3, 22, 27, 32, 34, 37, 39, 40, 47]. Due to the procedure oriented (e.g. fork2 join) source level paradigm, the computational task graph has a series parallel structure, which permits an efficient prediction, typically implemented through a recursive reduction scheme, based on the existence of a homomorphism between ....

F. Andr'e, J-L. Pazat, and H. Thomas, "Pandore: A system to manage data distribution," in Proc. 4th ACM Int. Conf. on Supercomputing, June 1990, pp. 380--388.

.... In this section, we briefly describe, at an abstract level, the support for dynamic data distributions in the Vienna Fortran Compiler System (VFCS) More details of the compilation strategy used in VFCS are given in [7, 17] Some of the issues discussed here are also being handled in other systems [1, 2, 8, 14]. The features required to manage dynamic data distribution comprise both compile time and run time elements. Most of these features are actually required to handle other aspects of Vienna Fortran: in particular, many of the problems posed by run time redistribution of data structures are the ....

F. Andr'e, J.-L. Pazat, and H. Thomas. PANDORE: A system to manage data distribution. In International Conference on Supercomputing, pp. 380--388, June 1990.

....of elements in x (the sum of the sizes of all the blocks, without ghost cells) 3. Create an array of 3 D decompositions, D3[num blocks] Each element of D3 corresponds to one block, and each decomposition is the same size in every dimension as its corresponding block. 4. Embed decomposition D3[1] into D1 at position 1, and all other decompositions D3[i] into D1 after D3[i Gamma 1] i.e. D3[i] starts right after D3[i Gamma 1] ends) 5. Distribute each decomposition in D3 (e.g. block wise in each of its dimensions) 6. Align each block in x with its corresponding decomposition in D3 ....

....provides facilities that support dynamic load balancing. DecTool [12] is an interactive environment designed to provide facilities for either automatic or manual decompositions of 2 D or 3 D discrete domains. There are a variety of compiler projects targeted at distributed memory multiprocessors [1, 9, 10, 14, 17, 18, 23, 25, 26, 27, 34, 35, 36, 40, 45]. Runtime compilation methods are employed in four of these projects; the Fortran D project [21] the Kali project [23] Marina Chen s work at Yale [30] and our PARTI project [32, 36, 37] The Kali compiler was the first compiler to implement inspector executor type runtime preprocessing [23] ....

F. Andr'e, J.-L. Pazat, and H. Thomas. PANDORE: A system to manage data distribution. In International Conference on Supercomputing, pages 380--388, June 1990.

....distributing iterations of loops are provided. In the Cray programming model, many of the features of shared memory parallel languages have been retained: these include critical sections, events and locks. New instructions for node I O are provided. Other systems include Pandore, a C based system [2], Id Nouveau, a compiler for a functional language [41] Oxygen [42] ASPAR [27] Adapt, developed at the University of Southampton [36] and the Yale Extensions [11] In a few systems, dynamic data distributions have been implemented within narrow constraints [3, 2] The systems described above ....

.... include Pandore, a C based system [2] Id Nouveau, a compiler for a functional language [41] Oxygen [42] ASPAR [27] Adapt, developed at the University of Southampton [36] and the Yale Extensions [11] In a few systems, dynamic data distributions have been implemented within narrow constraints [3, 2]. The systems described above are not the only efforts to provide either suitable language constructs for mapping code onto DMMPs or to generate message passing programs from higher level code. Other important approaches include Linda [1] Strand [15] and Booster [38] Knowledge based ....

F. Andr'e, J.-L. Pazat, and H. Thomas. PANDORE: A system to manage data distribution. In International Conference on Supercomputing, pages 380--388, June 1990.

.... language programming, they fall into the following categories: ffl Automatic parallelisation of conventional sequential languages, especially Fortran 77 and C, with processor topology and data mapping specified interactively or by program annotations (e.g. Superb [9] Parascope [11] and Pandore C [12]) ffl Extensions to conventional languages to explicitly express data parallel execution as well as the processor topology and data mapping, but without explicit message passing. Examples include Kali [13] Fortran D [5] and C [14] ffl New languages designed for efficient transformation to ....

F. Andr'e, J.-L. Pazat and H. Thomas. PANDORE: a system to manage data distribution, Proc. Int. Conf. on Supercomputing, 380-88, June 1990.

....distributed memory compiler that integrates and extends analysis and optimization techniques from many other research projects. It is related to other distributed memory compilation systems such as Al [33] CM Fortran [32] C [29] Dataparallel C [13] Dino [30] MIMDizer [15] Pandore [2], Paragon [9] and Spot [31] but mostly builds on the following research projects. Superb is a semi automatic parallelization tool that supports arbitrary user specified contiguous block distributions [12, 37] It originated overlaps as a means to both specify and store nonlocal data accesses. ....

F. Andr'e, J. Pazat, and H. Thomas. Pandore: A system to manage data distribution. In Proceedings of the 1990 ACM International Conference on Supercomputing, Amsterdam, The Netherlands, June 1990.

....will allocate a piece of memory on all the processors. Finally, the symbol can be used to specify a dynamic memory replication along a particular dimension, as in locate heap onto [ The following program creates a distributed linked list: template T[ distribute T[CYCLIC (64) int host[1]; align host[I] onto T[I] local int scanf(const char , int i = 0; struct cell int val; struct cell next; end, begin; locate heap onto T[0] begin = struct cell ) malloc (sizeof (struct cell) end = begin; host[0] scanf ( d , end val) while (end val = EOF) locate heap onto ....

....the language design and compiler levels. The above features in HPC are by no mean complete and we are actively working on further extensions. 6 Conclusion HPC follows the current trends of highperformance computing using C. It is related to other languages such as C [9] Hyper C [8] PANDORE [1], and pC [5] However, HPC is closer to HPF [3, 6] although it goes beyond HPF and, generally, the data parallel paradigm by providing support for dynamic data structures as they are available in C. The HPC language is currently under development at CRIM as part of the EPPP project. The ....

F. Andr'e, J-L. Pazat, and H. Thomas. PANDORE: A System to Manage Data Distribution. In J. Saltz and P. Mehrotra, editors, Languages, Compilers and Runtime Environments for Distributed Memory Machines, pages 177--184. Elsevier Science, 1992. ISBN: 0-444-88712-1.

.... can detect and exploit parallelism automatically, without requiring the user to specify single assignment (Crystal [29] Id Nouveau [32] all parallel loops (Al [36] Arf [37] Kali [26] parallel functions (C Dataparallel C [17, 33] Dino [34] parallel code blocks (Oxygen [35] Pandore [2]) or parallel array operations (CM Fortran [7] Paragon [10] The Fortran D compiler is similar to Callahan Kennedy [9] and Superb [15, 38] but applies analysis and optimization up front before code generation, rather than inserting guards and element wise messages then optimizing via program ....

F. Andr'e, J. Pazat, and H. Thomas. Pandore: A system to manage data distribution. In Proceedings of the 1990 ACM International Conference on Supercomputing, Amsterdam, The Netherlands, June 1990.

....At the end of the fan in, on all processors, the entries of nmsgs are identical. The value nmsgs(j) is equal to the number of messages that processor P j must send during the exchange phase. In our example scenario, we see that at the end of the fan in, the value of nmsgs on each processor will be [2,1,0] (Figure 5) Thus P 1 is able to determine that it needs to send data to two other (as yet unspecified) processors, P 2 needs to send data to one processor and P 3 does not need to send any data. At this point, each processor transmits to the appropriate processor, a list of required array ....

....In addition, it can be used by programmers in a wide range of applications. By contrast, programming environments such as those described by Baden and Williams are highly customized for use in specific application areas. There are a variety of compilers targeting distributed memory multiprocessors [44, 8, 33, 31, 1, 39]. With the exception of the Kali project [22] and the PARTI work described here and in [36, 29, 37] these compilers do not attempt to deal with loops having irregular references efficiently. The work described in this paper is also related to schemes to carry out distributed memory runtime ....

F. Andr' e, J.-L. Pazat, and H. Thomas, PANDORE: A system to manage data distribution, in International Conference on Supercomputing, June 1990, pp. 380--388.

....boundaries including inheriting the distribution of the actual arguments. Distribution inquiry functions facilitate the writing of library functions which are optimal for multiple incoming distributions. High Performance Fortran effort has been based on the above and other related projects [8, 27, 38, 48, 58, 59, 60]. In the next few sub sections we provide, short introduction to HPF concentrating on the features which are critical to parallel performance. 4.1 HPF Overview High Performance Fortran z is a set of extensions for Fortran 90 designed to allow specification of data parallel algorithms. The ....

....initial implementations in the near future. However, several research projects have built prototype compilers for HPF like languages. This includes the Kali compiler [33] the SUPERB project [67] on which the Vienna Fortran compiler is based, the Fortran D compiler [29] and several other efforts [8, 24, 27, 32, 38, 48, 58, 59, 60] that have contributed to the overall goal of compiling global name space programs for distributed memory SIMD and MIMD machines. pC is just one example of a number of efforts to add parallelism to C . While pC has been ported to a wide variety of machines including the TMC CM 5, Intel ....

F. Andr'e, J.-L. Pazat, and H. Thomas. PANDORE: A system to manage data distribution. In International Conference on Supercomputing, pages 380--388, June 1990.

....D compiler as a second generation distributed memory compiler that integrates and extends analysis and optimization techniques from many other research projects. It is related to other distributed memory compilation systems such as Al [30] CM Fortran [1] Dino [28] MIMDizer [15] Pandore [3], Paragon [10] and Spot [29] but mostly builds on the following research projects. SUPERB is a semi automatic parallelization tool that supports arbitrary user specified contiguous BLOCK distributions [13, 34] It originated overlaps as a means to both specify and store nonlocal data ....

F. Andr'e, J. Pazat, and H. Thomas. Pandore: A system to manage data distribution. In Proceedings of the 1990 ACM International Conference on Supercomputing, Amsterdam, The Netherlands, June 1990.

....is allowed to update this part. This mechanism is commonly referred to as the Exec mechanism in the community of data parallel computing. Similarly, the Refresh mechanism ensures that remote accesses are properly dealt with. Both mechanisms have been introduced in [4] and described formally in [3]. The EPEE toolbox provides various facilities for implementing these mechanisms, as illustrated in the following sections with the implementation of distributed matrices. 2.4 Matrices and Vectors in Paladin Paladin is built around the specifications of the basic entities of linear algebra: ....

F. Andr'e, J.L. Pazat, and H. Thomas. -- Pandore: a system to manage data distribution. -- In ACM International Conference on Supercomputing, June 11-15 1990.

....many of the features of shared memory parallel languages have been retained: these include critical sections, events and locks. New instructions for node I O are provided. Other systems include AL, which has been implemented on the Warp systolic array processor [40] Pandore, a C based system [2], Id Nouveau, a compiler for a functional language [33] Oxygen [36] ASPAR [22] Adapt, developed at the University of Southampton [29] and the Yale Extensions [10] In a few systems, dynamic data distributions have been implemented within narrow constraints [3, 2] The systems described above ....

.... [40] Pandore, a C based system [2] Id Nouveau, a compiler for a functional language [33] Oxygen [36] ASPAR [22] Adapt, developed at the University of Southampton [29] and the Yale Extensions [10] In a few systems, dynamic data distributions have been implemented within narrow constraints [3, 2]. The systems described above are not the only efforts to provide either suitable language constructs for mapping code onto DMMPs or to generate message passing programs from higher level code. Other important approaches include Linda [1] Strand [12] and Booster [31] 8 Conclusion In this ....

F. Andr'e, J.-L. Pazat, and H. Thomas. PANDORE: A system to manage data distribution. In International Conference on Supercomputing, pages 380--388, June 1990.

No context found.

F. Andr'e, J.-L. Pazat, and H. Thomas. PANDORE: A system to manage data distribution. In International Conference on Supercomputing, pages 380--388, Amsterdam, The Netherlands, June 1990.

First 50 documents

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