G.C. Fox et al., "Fortran D Language Specifications," Technical Report SCCS 42c, NPAC, Syracuse University, Syracuse, NY.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....expensive dot products, to complex heuristics. Specifying a data parallel computation in terms of a single element is the approach we have used in creating our data parallel extensions. We call this approach elementcentered. Fundamentally, this concept is not new to data parallel languages [2, 4, 6, 13, 15, 16, 17, 20, 21]. However, we have extended the notion to encompass subset level data parallelism. By subset level data parallelism we mean allowing the definition of operations in which subsets (as opposed to elements) are the data granules, e.g. a row or a column. This is not the same as applying an ....

....implementation executing concurrently with other mentat object member functions. Figure 6 illustrates both task and data parallel invocations within the same code fragment. 5: Related Work In terms of purely data parallel languages, Dataparallel C [13, 20, 21] pC [2, 16] C [15] Fortran D [6], Fortran 90 [1] and High Performance Fortran (HPF) 17] are the languages from which are related to our work. C and pC are based on C . Dataparallel C is based on C, but uses some ideas from object oriented language design. HPF s origin is Fortran. Our work differs from previous work in ....

G.C. Fox et al., "Fortran D Language Specifications," Technical Report SCCS 42c, NPAC, Syracuse University, Syracuse, NY.

.... been proposed in many systems by various researchers Configuration file: Input: P: MAIN inf P:MVM Output: P: XC xf P: MAIN P:XC inf P: MVM ouf P:XC mf MAIN t xf I I MVM ouf xc Figure 3: Configuration for an iterative solver in a) textual and b) graphical form [11, 12, 13, 14, 15] and are used mainly as compiler directives. In our approach annotations limit the feasible mappings of computation onto the processors. Hence, they are used only during the decomposition of a process into smaller fragments. This kind of annotation is similar to 0N clause as used in the Kali ....

....as compiler directives. In our approach annotations limit the feasible mappings of computation onto the processors. Hence, they are used only during the decomposition of a process into smaller fragments. This kind of annotation is similar to 0N clause as used in the Kali compiler [11] Fortran D [12] or Vienna Fortran [13] Annotation does not have any effect on the result computed by a program. Consequently, sequential programs that have manifested their correctness over many years of usage are good candidates for parallelization through annotations. Being orthogonal to the program ....

[Article contains additional citation context not shown here]

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C. Tseng, and W. Wu, "Fortran D language specification," Tech. Rep. COMP TR90079, Department of Computer Science, Rice University, Houston, March 1991.

....array it will not be allocated on the Neuor4 and all the parallel sections consisting of it will not be parallelized. The array distribution unit determines how the arrays will be distributed: block, cyclic, or block cyclic. The theory behind different data distribution approaches can be found in [9]. The initial version of the SIMDizer will distribute all the arrays cyclically. Fortunately the Neuor4 code generator, by making use of the index register Neuro4 chip, performs the matrix staggering, explained in pp. 181 183 of [10] in software for high performance of matrix multiplication ....

G. Fox et. al., "Fortran D Language Specification", Technical report TR90-141, Computer Science Department, Rice University, December 1990.

....under development at Rice University are examples of projects in this area. Other groups have targetted their effort to the effective compilation of programs containing the specification of the data distribution, such as the VFCS [5] for the Vienna Fortran language [22] the Fortran D compiler [10, 21]. All the automatic approaches to automatically distribute data perform the job in two main independent steps: alignment and distribution. The alignment step tries to find appropriate alignments between all arrays in a block of code, that is, to decide for each array the dimensions that will be ....

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C. Tseng and M. Wu, "Fortran D Language Specification". Technical Report CRPC TR 90-141, Department of Computer Science, Rice University, December 1990.

....data alignment deals with how data arrays should be aligned with respect to one another. The purpose of data decomposition is to balance the computational load and minimize the communication overheads. Many data parallel programming languages, such as High Performance Fortran (HPF) 8] Fortran D [5], Vienna Fortran [33] and High Performance C (HPC) 28] provide compiler directives for programmers to specify array distribution. The array distribution provided by those languages, in general, can be classified into two categories, regular and irregular. The regular array distribution, in ....

....there are two source sections in SLA 1 [1 : BC] of a source processor P 1 . The destination distribution pattern position of SLA 1 [1] A[4] is equal to three. Since array elements in a source section have consecutive global array indices, from Definition 6, we can derive that DDPP(SLA 1 [2] A[5]) 0 and DDPP(SLA 1 [3] A[6] 1, respectively. The corresponding destination processors of SLA 1 [1] SLA 1 [2] and SLA 1 [3] are equal to DDPP SLA t P ( 1 1 1 = DDPP SLA t P ( 1 2 0 = and DDPP SLA t P ( 1 3 0 = respectively, where t = 2. For a source processor P ....

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C.-W. Tseng, and M. Wu, "Fortran-D Language Specification," Technical Report TR-91-170, Dept. of Computer Science, Rice Univ., Dec. 1991.

....one is to give him the responsability of deciding how data structures have to be aligned each other, and distributed among the memory modules of the system. In this case, the sequential code is annotated with directives and executable statements, offered by current languages such as 2 Fortran D [FHKK90], High Performance Fortran [HPFF93] or Vienna Fortran [ZBCM91] The second alternative is to leave this task to the compiler. In any case, classical aspects such as data movement, parallelism, and load balance have to be taken into consideration in a unified way to efficiently solve the data ....

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C. Tseng and M. Wu, "Fortran D Language Specification". Technical Report CRPC TR 90-141, Department of Computer Science, Rice University, December 1990.

.... topology( mesh ,M,N, i M, j N return(M i j) default return( 1) Data distribution is supported in the form of blocked distributions of arrays of singleassignment variables. In a blocked distribution, each processor is allocated a contiguous block of array elements [18]. Elements of a distributed array are accessed in the same manner as ordinary array elements. More complex distributions and data structures, such as those supported in data parallel programming languages [30, 18] can be integrated in the same manner, but are not supported in the current PCN ....

....a blocked distribution, each processor is allocated a contiguous block of array elements [18] Elements of a distributed array are accessed in the same manner as ordinary array elements. More complex distributions and data structures, such as those supported in data parallel programming languages [30, 18], can be integrated in the same manner, but are not supported in the current PCN compiler. A keyword port is used to declare these distributed arrays. For example, the following procedure declares a distributed array P of single assignment variables with one element on each node of the current ....

[Article contains additional citation context not shown here]

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C. Tseng, and M. Wu, "Fortran D language specification," Technical Report TR90-141, Computer Science, Rice Univ., Houston, Texas, 1990. 14

....and task execution in the case of nontrivial and or dynamic data dependencies. But making the domain decomposition explicit is not always sufficient. HPF [33] and the related projects HPC [32] Vienna Fortran [14] C [35, pp. 450 459] Annai [16] CM Fortran [53] PC Sage [23] Fortran D [22] (augmented with the CHAOS runtime support procedures [40] Mentat [24] the explicit parallel Fortran syntax bindings from the draft X3H5 document, etc. allow the user to specify or to suggest how to distribute the data. But they do not provide control structures to express parallelism ....

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C. Tseng, M. Wu, "Fortran D language specification," Technical Report TR90079, Dept. of Computer Science, Rice University, December 1990

....of repetitive operations. The programming model is SPMD (Single Program MultipleData) for the nCUBE 2 and SIMD (Single Instruction Multiple Data) for the MasPar MP 1. A number of data distribution (or machine mapping) strategies are possible. Using terminology from the Fortran D specification [13], the results presented here are from three strategies: 1) DISTRIBUTE (CYCLIC, CYCLIC) 2) DISTRIBUTE (BLOCK, and (3) DISTRIBUTE (CYCLIC, These are shown in Figure 1. All are two dimensional distributions of the matrix among processors, where the latter two divide along the first dimension ....

Fox, G., et. al., "Fortran D Language Specification," Tech. Report CRPC-TR90079, Rice Univ., Dec. 1990.

....of the computation are mapped, and these parts usually have a clear correspondence with the modules we discussed above. In data parallelism data is mapped and, by implication computation follows. An exapple of this approach is the work on Fortran language extensions and related areas by Fox et al. 1992] and Benkner et al. 1992] Here data arrays are mapped with compiler directives and computation follows implictly by the Owner Computes rule whereby the computation required to assign to a data array element is performed by the processor to which that data element has been mapped. We can ....

Fox, G., Hiranandani, S., Kennedy, K., Koelbel, C., Kremer, U., Tseng, C.-W., and Wu, M.- Y. (1992). `Fortran D Language Specification'. Technical report, Center for Research on Parallel Computations, Rice University, Houston, TX-77251, USA.

....case of distributed memory machines, consideration is given to the problems of data partitioning and distribution, and work allocation in order to maximize accesses to local memory. The compiler support for loop parallelism and data distribution is based on features of Cedar Fortran [3] Fortran D [4], Cray MPP Fortran [5] and HPF [6] In both the centralized and distributed memory cases, an autoscheduling program can execute on a physical partition of a multiprocessor. This partition is not necessarily fixed, but may vary throughout the execution of the program. The level of exploited ....

....present, are executed once for each processor that participated in the execution of the loop after it has executed all of its iterations. They can be used to save temporary values that each processor computed. CDOALL and CDOACROSS loops do not allow ENDLOOP statements. 3.1. 2 Fortran D Fortran D [4, 21, 22, 23, 24] is a version of Fortran enhanced with data decomposition specifications. Fortran D is targeted at data parallel numerical applications, and is primarily concerned with the exploitation of fine grain parallelism at the array operation level. The set 17 of Fortran D extensions is compatible with ....

[Article contains additional citation context not shown here]

G. Fox, S. Hiranandani, K. Kennedy, C. Koebel, U. Kremer, C.-W. Tseng, and M.-Y. Wu, "Fortran D language specification," Tech. Rep. COMP TR90-141, Department of Computer Science, Rice University, December 1990.

....the data parallel or SPMD programming model based on a single name space provides an easier and familiar programming style for users. In order to provide high level language support for data parallel programming, several data parallel Fortran languages have been proposed, such as Fortran D [1, 2] and Vienna Fortran [3] In an effort to standardize data parallel Fortran programming, HPF (High Performance Fortran) 4] an extension of Fortran 90, is being proposed as a standard for distributed memory machines. An essential part of these data parallel Fortran languages is the specification ....

....process configurations. The calling environment creates and fills the appropriate data structures as dictated by the interface introduced in Section 2. The reorder parameter is used to associate the two Cartesian topologies since they have different shape. dimsize = 3; srccomm[0] 5; srccomm[1] = 4; srccomm[2] 2; destcomm[0] 4; destcomm[1] 1; destcomm[2] 10; Set periodicity to FALSE for(i=0;i dimsize;i ) periods[i] 0; Construct a cartesian topology for the source processor configuration; allow reordering MPIMakecart ....

[Article contains additional citation context not shown here]

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C.-W. Tseng, and M.-Y. Wu, "Fortran D language specification," Tech. Rep. COMP TR90-141, Rice University, Department of Computer Science, Dec. 1990.

....irregular data distribution (e.g. a CFD problem defined on a nonrectangular grid) but the data are defined at compile time and the dependency is data parallel. Such irregular problems has been dealt with data parallel languages supporting irregular data distributions. Examples are Fortran D [3] and Vienna Fortran [1] The second class comprises problems whose data access patterns and workload are only known at run time. This complicates the task of explicitly specify the data distribution (e.g. a mapping of array elements to processors) in a data parallel language. One of the solutions ....

G. Fox et al., "Fortran D language specification", Tech. Report CRPC-TR90097, Center for Research on Parallel Computation, Rice Univ., Houston, Texas, 1990.

....questions will be part of the research effort. 5.1.2. Data Parallel Translations In the data parallel model, parallelism is obtained by performing the same set of operations on many data elements simultaneously. A language can support data parallelism either implicitly or explicitly. FORTRAN D [8] is an example of an implicit data parallel language. In FORTRAN D the programmer specifies how the data, in the form of arrays, is to be distributed. Inner loop iterations that act on the data are executed in parallel. The data parallelism is implicit because the programmer simply writes ....

....they are all ad hoc. VMPP has the advantage that it is a general solution and also provides a better solution in most cases. 6.1. Single Language Solutions to the Portability Problem There are currently several parallel languages designed to provide portability across parallel architectures [8, 14]. The approach taken by these languages is to provide a single parallel programming language along with compilers for a variety of parallel architectures. The language is usually designed to be relatively architecture independent. The single language approach suffers from two problems. First, ....

Geoffrey Fox, Seema Hiranandani, Ken Kennedy, Charles Koelbel, Uli Kremer, Chau-Wen Tseng, and Min-You Wu, "FORTRAN D Language Specification", Technical Report TR90141, Department of Computer Science, Rice University, Houston Texas, 1990.

....one is to give him the responsability of deciding how data structures have to be aligned each other, and distributed among the memory modules of the system. In this case, the sequential code is annotated with directives and executable statements, offered by current languages such as Fortran D [FHKK90], High Performance Fortran [HPFF93] or Vienna Fortran [ZBCM91] The second alternative is to leave this task to the compiler. In any case, classical aspects such as data movement, parallelism, and load balance have to be taken into consideration in a unified way to efficiently solve the data ....

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C. Tseng and M. Wu, "Fortran D Language Specification". Technical Report CRPC TR 90-141, Department of Computer Science, Rice University, December 1990.

....hold the promise of scalable performance. However they have been difficult to program and recent effort has concentrated on exploiting data parallelism within FORTRAN programs by providing user inserted data distribution pragmas. Notable examples are High Performance FORTRAN (HPF) FORTRAN D [4] and Vienna FORTRAN [1] The scheduling of data and work to processors can have a dramatic effect on program performance. Furthermore, there are conflicting factors which have to be resolved for a successful implementation. It is for this reason that many researchers have suggested that the ....

Fox G., Hiranandani S., Kennedy K., Koelbel C., Kremer U., Tseng C-W. and Wu MY. , "FORTRAN D Language Specification", Rice COMP TR90-141, Department of Computer Science, Rice University, February 1991.

....problem or a particular machine. The specificity of these models may be quite valuable to certain users but of little value to others. Much effort has been expended on the development of data parallel languages intended to free the user of responsibility of managing communication primitives, e.g. [9, 20, 30]. However, these languages still require that the user choose the number of processors and specify how to distribute the data among the processors, or accept default distributions. Because of architectural differences among various parallel machines and differences in compilers, a program ....

....provide mechanisms, typically compiler directives, by which the programmer may distribute data. For specificity, we use the terminology and data distribution model of High Performance Fortran (HPF) in the work below, although other parallel languages such as Vienna Fortran [30] Fortran D [9] and CRAFT (Cray Adaptive Fortran) provide similar functionality. The HPF data distribution model is illustrated in Figure 1, adapted from [17] The ALIGN and REALIGN directives are provided to specify that the distribution of one array should depend on the distribution of another array. The ....

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C. Tseng, and M. Wu, "Fortran D language specification," Department of Computer Science Rice COMP TR90079, Rice University, March 1991.

....The recent development of the High Performance Fortran (HPF) 1] standard has given hope to the scientific computing community that the power of parallel computers may soon be within grasp. The initial work on HPF, based on contributions from a number of Fortran dialects such as Fortran D [2], Connection Machine Fortran [3] and Vienna Fortran [4] has primarily focused on compile time optimizations for regular, static applications. An important class of scientific calculations employ dynamic, irregular, block This work was supported by NSF contract ASC 9110793 and ONR contract ....

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C. Tseng, and M. Wu, "Fortran d language specification," Tech. Rep. TR90-141, Dept. of Computer Science, Rice University, Houston, TX, December 1989.

....was supported in part by the NSF grants CDA9121641 and MIP 9204066. or any computer with non uniform memory access time) several languages have been developed that enable the programmer to include information concerning the way in which data should be distributed among the various processors [1, 2]. If a proper alignment of data is chosen, then communication costs may be minimized. In this paper, we shall use Fortran D language extension proposed by Rice University for the purpose of illustration. In Fortran D, the DECOMPOSITION statement indicates how an array is to be used as a frame of ....

....of reference, or template. The ALIGN statement indicates how an array should be aligned with the template. The DISTRIBUTE statement indicates how the template should be distributed to different processors. For a more complete explanation of the syntax of this and other Fortran D instructions, see [1]. Li and Chen have proven that the general problem of data distribution on distributed memory machines is NP complete [3] However, even to find a good, not necessary the best, distribution for the arrays of a complex program on a distributed memory parallel computer may be a formidable task if ....

[Article contains additional citation context not shown here]

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C.-W. Tseng, and M.-Y. Wu, "Fortran D language specification," Tech. Rep. COMP TR90-141, Rice University, Department of Computer Science, Dec. 1990.

....in use but many applications require no format at all. These matrix free methods effectively perform matrix vector products implicitly to avoid storage of the matrix. However, in the parallel computing context, efficiency issues (minimizing communication costs and maximizing computational rates)[4,8,14] and programming issues make such software critical to the development of advanced applications[15] In this paper, a set of programming tools are described to help transform a global matrix description to a distributed matrix partitioned over multiple processors such that a general unstructured ....

....the vector components among the processors. Additionally, the user needs to give the sparsity pattern of the global matrix where each processor specifies only the sparsity pattern for its rows. The partitioning of the vector elements can be done in a linear fashion or so called BLOCK distribution [4](i.e. elements 0 to N P 1 are assigned to processor 0, elements N P to 2N P 1 are assigned to processor 1, etc. where N is the total number of vector components and P is the total number of processors) However, there are much better methods such as those in Chaco[5] which perform this task and ....

G. C. Fox, S. Hiranandani, K. Kennedy, C.Koelbel, U. Kremer, C. Tseng, M. Wu, "FORTRAN D Language Specification", CRPC Technical Report CRPC-TR9009, April 1991

....Data) mode of computation, in which the same program is executed on different processors with different data. Specifically, multicast is fundamental to several operations, such as replication and barrier synchronization [3] supported in data parallel languages, including Fortran 90 and others [4, 5]. Third, if a distributed shared memory paradigm is supported, then multicast services may be used to efficiently support shared data invalidation and updating [6] Most existing MPCs support only unicast communication in hardware. In these environments, multicast must be implemented in software ....

....this time, contention will again occur for the (1; 1) to (1; 2) channel. The multicast tree in Figure 4(e) which is based on the methods presented in the following sections, is contention free regardless of message length or receiving latency. 1,1 1,2 1,3 2,0 1,0 0,0 3,2 2,2 [1] 2] 3] 2] 3] [4] [3] 4] a) A multicast tree based on store and forward switching 1,1 1,2 1,3 2,0 1,0 0,0 3,2 2,2 [1] 2] 3] 2] 3] 4] 3] b) A multicast tree based on wormhole routing [3] 2,0 1,0 1,1 1,2 1,0 0,0 1,1 2,1 3,1 3,2 1,1 1,2 1,3 [1] 2] 3] 3] 2] 3] 3] c) Collision occurs at the channel ....

[Article contains additional citation context not shown here]

G. Fox, S. Hiranandani, , K. Kennedy, C. Koelbel, U. Kremer, C.-W. Tseng, and M.-Y. Wu, "Fortran D language specification," Tech. Rep. TR90-141, Department of Computer Science, Rice University, Dec. 1990.

....information can be obtained for an individual AAU, cumulatively for a branch of the AAG (i.e. sub AAG) or for the entire AAG. 3 A HPF Fortran 90D Performance Prediction Framework 3. 1 An Overview of HPF Fortran 90D High Performance Fortran (HPF) 2] is based on the research language Fortran 90D [3] and provides a minimal set of extensions to Fortran 90 to support the data parallel programming model. Extensions incorporated into HPF Fortran 90D provide a means for explicit expression of parallelism and data mapping. These extensions include compiler directives which are used to advise the ....

Geoffrey C. Fox, Seema Hiranandani, Ken Kennedy, Charles Koebel, Uli Kremer, Chau-Wen Tseng, and Min-You Wu, "Fortran D Language Specifications", Technical Report SCCS 42c, Northeast Parallel Architectures Center, Syracuse University, Syracuse NY 13244-4100, Dec. 1990, Available via WWW at http://www.npac.syr.edu.

....the data parallel or SPMD programming model based on a single name space provides an easier and familiar programming style for users. In order to provide high level language support for dataparallel programming, several data parallel Fortran extensions have been proposed, such as Fortran D [1] and Vienna Fortran [2] In an effort to standardize data parallel Fortran programming, HPF (High Performance Fortran) 3] is being proposed as a standard by the High Performance Fortran Forum led by Rice University for distributed memory machines. In this paper, we shall use HPF to illustrate the ....

.... Theta Theta Theta Theta (cyclic, Figure 8. Absolute execution time for implementations (cyclic, and (block, 256 512 768 1024 Problem Size (n) 1 8 16 24 32 40 48 56 64 [ 16,44)U(44,64) 16,64] 16,64] 8,32)U(32,64) 8,64] 8,64] 2,27)U(27,64) 2,32] 2,64] 1,14)U(14,32) [1,16] [1,32] Number of Processors (p) block, cyclic, combined Figure 9. Various ranges of scalability of Gaussian elimination algorithm for different problem sizes measured on a 64 node nCUBE 2. 1 8 16 24 32 40 48 56 64 Implementation (cyclic, implementation (block, Beyond Scalability Range ....

[Article contains additional citation context not shown here]

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C.-W. Tseng, and M.-Y. Wu, "Fortran D language specification," Tech. Rep. COMP TR90-141, Rice University, Department of Computer Science, Dec. 1990.

....as they develop, rather than at compile time. This facilitates larger grain computation by eliminating many small grain control actors which are inefficient in a distributed memory environment. Second, Mentat supports both task and data parallelism, not just data parallelism as in [8] 15][18][28] 33] Third, Mentat operates over a spectrum of architectures, from loosely coupled heterogeneous networks of workstations, to tightly coupled multicomputers. Finally, Mentat has a scalable, distributed, control mechanism. This includes both the mechanism used to construct and modify program ....

....performance via parallelism rather than distributed execution. 5.2 Compiled Distributed Memory Systems Until recently, there were few results for compiled as opposed to hand coded applications on distributed memory machines. There are now several active projects in this area, Fortran D [18], HP Fortran [33] Dataparallel C [37] Paragon [15] and the Inspector executor [48] model to name a few. These are primarily data parallel languages with [48] tied to a data parallel model of computation. What differentiates our work from theirs is that Mentat exploits opportunities for both ....

[Article contains additional citation context not shown here]

G. C. Fox, et al., "Fortran D Language Specifications," Technical Report SCCS 42c, NPAC, Syracuse University, Syracuse, NY.

....The time complexity to construct the whole class table is O(s 2 ) Keywords: Communication set, data parallel language, distributed memory multicomputers, HPF, parallelizing compilers, SPMD. 1 Introduction Data parallel languages, such as High Performance Fortran(HPF) 1] and Fortran D [3, 7], support the global name spaces and provide the directives for programmers to specify distributions of arrays at language level. For a data parallel program in which the distribution could be specified by the programmers or by a separate compilation phase, compiler must automatically partition ....

....an assignment statement will be specified a distribution with directives provided by data parallel languages. We now introduce common distributions provided by general data parallel languages. 2. 2 Data Distribution Data parallel languages, such as High Performance Fortran(HPF) 1] and Fortran D [3], provide regular distributions including block, cyclic, block cyclic distributions. While the array elements are distributed over processors, we need to store these array elements in local memory. The index of an array element in original program is called a global index and the index of an array ....

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C.-W. Tseng, and M. Wu, "Fortran-D Language Specification," Technical Report TR-91-170, Dept. of Computer Science, Rice University, Dec. 1991.

....making using only local information. Language Extensions and Run Time Libraries In [FiBa95] language extensions are proposed to support irregular block decompositions of space such as Orthogonal Recursive Bisection in a high level language such as High Performance FORTRAN [HPF] or FORTRAN D [FTND]. A number of run time libraries focused on solving similar problems have been developed including: LPARX University of California at San Diego Global Arrays (GA) Pacific Northwest Laboratory [NiHaLi94] RSL Runtime System Library Argonne National Laboratory BlockComm Argonne National ....

Fox G, Hiranandani S, Kennedy K, Koelbel C, Kremer U, Tseng C, and Wu M, "Fortran D Language Specification", Technical Report TR90-141, Dept. of Computer Science, Rice University, Houston, TX, December 1989.

....perspective. Users can exploit any combination of either set of features without ambiguity, and thus are free to define a program from whatever perspective is most appropriate to the problem at hand. The data distribution declarations were adapted in part from Rice University s Fortran D project [4] and Vienna Fortran[7] The work distribution directives are adapted from a wide variety of sources, including Cray Autotasking(TM) see [1] and [3] and workshops on parallel programming held at the University of Illinois. Concepts embraced in this model can be found in many other sources as ....

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C.-W. Tseng, and M.-Y. Wu, "Fortran D Language Specification," Rice University, Houston, Texas, 1991.

....matrix algebra by using the physical communication layer and the already existing single node support in terms of the field objects, now playing the role of node sections of the domain decomposed global fields. Such construction represents the run time interpreted version of the Fortran90D model [FHKKKTW90]. Compiler directives are replaced by interpreter directives , i.e. MovieScript tools for data decomposition which can be employed in the dynamic realtime mode. Various interface models to the compiled Fortran90D environment can be also constructed. Furthermore, since arithmetic doesn t play any ....

Fox, G.C., Hiranadani, S., Kennedy, K., Koelbel, Ch., Kremer, U., Chau-Wen Tseng and Min-You Wu, "Fortran D Language Specification", SCCS/42C, 1990.

....irregular data distribution (e.g. a CFD problem defined on a nonrectangular grid) but the data are defined at compile time and the dependency is data parallel. Such irregular problems has been dealt with data parallel languages supporting irregular data distributions. Examples are Fortran D [3] and Vienna Fortran [1] The second class comprises problems whose data access patterns and workload are only known at run time. This complicates the task of explicitly specify the data distribution (e.g. a mapping of array elements to processors) in a data parallel languange. One of the solutions ....

G. Fox et al., "Fortran D language specification", Tech. Report CRPC-TR90097, Center for Research on Parallel Computation, Rice Univ., Houston, Texas, 1990.

....recently, whereby a sophisticated compiler generates data movement operations from shared memory parallel programs. In order to support a portable and scalable software design across different platforms, several data parallel languages have been proposed, including Fortran 90, CRPC s Fortran D [2], Cray s MPP Fortran [3] CM 5 s Fortran [4] and High Performance Fortran. These languages support a variety of data movement and process control operations. Such operations can be classified as replication, reduction, segmented scan, and permutation. Data movement operations are often applied to ....

....transfer a message from a source node to a set of destination nodes. In order to facilitate use of the multicast algorithms described above, ComPaSS accommodates both static and dynamic groups. Static groups are those whose membership is known prior to run time. Dynamic Only Router Used [2] [3] 2] 3] 3] 11010 01110 01000 11100 11011 00001 01101 11110 11000 11101 11001 10001 00000 10100 10010 00110 00001 11011 10111 Destination Node Source Node 11010 i 01100 11001 d relative chain 11010 [1] Figure 3. An example U cube tree groups are those whose membership cannot be determined ....

[Article contains additional citation context not shown here]

G. Fox, S. Hiranandani, , K. Kennedy, C. Koelbel, U. Kremer, C.-W. Tseng, and M.-Y. Wu, "Fortran D language specification," Tech. Rep. TR90-141, Department of Computer Science, Rice University, Dec. 1990.

....contacting the other processor. Examples Cray T3D, Meiko CS 2, Fujitsu VPP500. 3. Single address space with cache coherency (SVM) Write ownership of data can pass from one processor to another. Examples KSR 1 and 2, Convex Exemplar. Of the eight dialects we survey here, HPF [6] Fortran D [5] and Vienna Fortran [12] are targeted at type 1 machines, Cray MPP Fortran [11] at type 2, and the remainder (KSR Fortran [7] Convex SPP Fortran [3] Fortran S [2] and SVM Fortran [1] at type 3. In these dialects we can distinguish three different forms of scheduling ffl Direct loop ....

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C-W. Tseng and M-Y Wu, "Fortran D language Specification" Rice University Dept. of Computer Science Technical Report TR90-141, 1990.

....strategy is to provide libraries of callbacks for well known computational structures and this technique is discussed in Section 2.5. Current compiler technology also looks promising for generating the SPMD task implementation and automatically decomposing the data domain for regular problems [2][9] 2 5 5 8 2 5 5 8 2 5 5 8 (a) 1 D (b) 2 D (c) tree Figure 6. Topology dependent partition map (numPDUs = 20) 13 2.4 Partitioning Partitioning is the second stage in the three stage process introduced in Section 1.0. Once processor availability has determined the available set of processors ....

G. Fox et al, "Fortran D Language Specification," TR90-141, Department of Computer Science, Rice University, December 1990.

.... to maximize the system performance by balancing the computational load and by minimizing remote memory accesses (or communication messages) In order to provide high level language support for data parallel programming, several dataparallel Fortran extensions have been proposed, such as Fortran D [1] and Vienna Fortran [2] In an effort to standardize data parallel Fortran programming, HPF (High Performance Fortran) is being proposed as a standard by the High Performance Fortran Forum led by Rice University [3] for distributed memory machines. An essential part of these data parallel Fortran ....

....of linear equations with Gaussian Elimination and Backward Substitution, and a Purdue set benchmark loop. Section 7 concludes the paper. 2 Data Distribution Specification BLOCK and CYCLIC distributions comprise two essential distributions, known as regular distributions, by HPF [3] Fortran D [1], and Vienna Fortran [2] With the BLOCK specification, contiguous, evenly sized segments of an array dimension are distributed to each processor. With the CYCLIC specification, elements of a dimension of an array are assigned to each processor in a round robin fashion. These distributions are ....

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C.-W. Tseng, and M.-Y. Wu, "Fortran D language specification," Tech. Rep. COMP TR90-141, Rice University, Department of Computer Science, Dec. 1990.

....matrix algebra by using the physical communication layer and the already existing single node support in terms of the field objects, now playing the role of node sections of the domain decomposed global fields. Such construction represents the run time interpreted version of the Fortran90D model [FHKKKTW90]. Compiler directives are replaced by interpreter directives , i.e. MovieScript tools for data decomposition which can be employed in the dynamic real time mode. Various interface models to the compiled Fortran90D environment can be also constructed. Furthermore, since arithmetic doesn t play any ....

Fox, G.C., Hiranadani, S., Kennedy, K., Koelbel, Ch., Kremer, U., Chau-Wen Tseng and Min-You Wu, "Fortran D Language Specification", SCCS/42C, 1990.

....in much of the HDCS work are interoperability, sharing, and availability. Unlike our work, high performance is not the objective. Applications portability across parallel architectures is an objective of many projects. Examples include PVM[28] Linda[7] the Argonne P4 macros[1] and Fortran D [11]. Our effort shares with these and other projects the basic idea of providing a portable virtual machine to the programmer. The primary difference is the level of the abstraction. Low level abstractions such as in [1,7,28] require the programmer to operate at the assembly language level of ....

....of providing a portable virtual machine to the programmer. The primary difference is the level of the abstraction. Low level abstractions such as in [1,7,28] require the programmer to operate at the assembly language level of parallelism. This makes writing parallel programs more difficult. Others [3,11] share our philosophy of providing a higher level language interface in order to simplify applications development. Our work differs from Fortran D [11] in several ways. First, Fortran D is portable but is not targeted to a heterogeneous environment. Second, Fortran D supports data parallelism ....

[Article contains additional citation context not shown here]

G. C. Fox, et al., "Fortran D Language Specifications," Technical Report SCCS 42c, NPAC, Syracuse University, Syracuse, NY.

....after Fortran P Translation. June 24, MPP 13 Fortran P Note that the CMF LAYOUT directive can be used to specify the layout of each array dimension as either parallel ( NEWS, across processor nodes) or serial ( SERIAL, within a single node) Other MPP languages have similar directives [Bla90] [FoH90] , HPF92] PaM92] This subroutine operates on 1 D arrays extracted from a 2 D or 3 D subdomain. Note that the array extent n must be adjusted to reflect the size of the 2 D subdomain allocated to each node. The parameters NODE X and NODE Y indicate the extent of the processor grid in the x and ....

G. Fox, S. Hiranandani, K. Kennedy, C. Koebel, U. Kremer, C. Tseng, M. Wu, "Fortran D Language Specification," Tech. Report TR 90-141, Dept. of Computer Science, Rice University, Dec. 1990.

....language extensions that permit the explicit representation of parallelism, where they are necessary for the construction of compilers that generate target specific source code containing calls to run time systems that support concurrency. Examples include the Fortran D and Fortran 90D compilers [1], and the pC compiler [5] Source tosource translation systems are also used to construct compilers that tackle the problem of discovering parallelism in sequential source code. These compilers apply restructuring transformations to generate a version of the program that uses directives or ....

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C. Tseng, and M. Wu. "Fortran D Language Specification." Tech. Report COMP TR90079 from the Dept. Computer Science, Rice University, March 1991.

....technology. Besides its traditional advantages, interpretation has also proved its value in cross platform executability and security [10] This development has already motivated some dedicated research in interpreters such as project Rocky being conducted in University of Washington [5]. 12 There is substantial research on optimizing Java both in academia and industry, including industry leaders such as Sun and Microsoft. Many issues are being explored, we note Optimizing Java compilers IBM is developing optimizing Java compiler that produces binary code for IBM machine ....

Fox, G.C., Hiranadani, S., Kennedy, K., Koelbel, C., Kremer, U., Tseng, C.W., Wu, M.Y., "FortranD Language Specifications", Rice COMP TR90079, December 1990.

....Ken Kennedy and then others, on parallel high level languages especially Fortran. This work was largely part of Fox s activity in CRPC (The Center of Research in Parallel Computation) which played a major role in the development of HPF with prototype compilers and community language definition [15]. This led a major project from ARPA to develop in collaboration with Rice a prototype High Performance Fortran (Fortran90D) compiler, including a High Performance Fortran interpreter that was demonstrated at Supercomputing 93, 6] 13] This early HPF compiler was licensed by the Portland Group ....

Fox, G.C., Hiranadani, S., Kennedy, K., Koelbel, C., Kremer, U., Tseng, C.W., Wu, M.Y., "FortranD Language Specifications", Rice COMP TR90079, December 1990.

....validate the approach and the framework. Section 6 discusses some related research. Finally, Section 7 presents some concluding remarks and discusses future extensions to the project. 2 An Overview of HPF Fortran 90D High Performance Fortran (HPF) 2] is based on the research language Fortran 90D [3] and provides a minimal set of extensions to Fortran 90 to support the data parallel programming model 2 . Extensions incorporated into HPF Fortran 90D provide a means for explicit expression of parallelism and data mapping. These extensions include compiler directives which are used to advice ....

Geoffrey C. Fox, Seema Hiranandani, Ken Kennedy, Charles Koebel, Uli Kremer, Chau-Wen Tseng, and Min-You Wu, "Fortran D Language Specifications", Technical Report SCCS 42c, Northeast Parallel Architectures Center, Syracuse University, Syracuse NY 13244-4100, Dec. 1990.

....programming or the development of new languages. To address some of the above issues, a specification for a language called High Performance Fortran (HPF) was adopted in 1993 [18] The HPF language specification is based in part on experience gained with the research languages such as Fortran D [21, 22, 10] and Vienna Fortran [5] HPF is based on Fortran 90 [1, 33] an extension of Fortran 77 that allows new features to be integrated into existing code in a controlled, evolutionary manner. Fortran 90 includes major new features (over Fortran 77) in a broad range of areas, including new control ....

Fox, G.C., Hiranadani, S., Kennedy, K., Koelbel, C., Kremmer, U., Tseng, C.W., and Wu, M., "Fortran D language specification," Technical Report, Rice and Syracuse Universities, 1992.

....a mechanism for specifying data placement. Unless the compiler can choose the right data placement automatically a formidable task this model will not be very useful on distributed memory machines. To address this problem, we have developed an extended version of Fortran, called Fortran D [75], which enables the programmer to explicitly specify data distribution and alignment on a multiprocessor system. Furthermore, the specification can be expressed in a machine independent form, making it possible for the programmer to write a single program image that can be compiled onto different ....

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C. Tseng, and M. Wu, "Fortran D language specification," Tech. Rep. TR90-141, Dept. of Computer Science, Rice University, Dec. 1990.

No context found.

Fox, G.C., Hiranadani, S., Kennedy, K., Koelbel, C., Kremmer, U., Tseng, C.W., and Wu, M., "Fortran D language specification," Technical Report, Rice and Syracuse Universities, 1992.

....not be designed for a particular machine model it expresses problem and not machine characteristics. Table 6: Candidate Software Paradigms for Each Problem Architectures ffl Synchronous: High Performance Fortran (HPF) Foster:95a] HPFCSep:95a] Koelbel:94a] Fortran 77D [Bozkus:93a] [Fox:91e], Hiranandani:92c] Vienna Fortran [Chapman:92b] C [Hatcher:91a;91b] Crystal [Chen:88b] APL; Fortran for SIMD parallel computers ffl Loosely Synchronous: Extensions of the above, especially HPF [Chapman:94b] Choudhary:92d] HPF:94a] and parallel C [Bodin:91a] Chandy:93a] ....

Fox, G. C., Hiranandani, S., Kennedy, K., Koelbel, C., Kremer, U., Tseng, C.-W., and Wu, M.-Y. "Fortran D language specification. " Technical Report SCCS-42c, Syracuse University, Syracuse, NY, April 1991. Rice Center for Research in Parallel Computation; CRPCTR90079.

....development of ESPial and outlines the structure and operation on its graphical user interface. Finally Section 6 presents some concluding remarks and discusses future extensions. 2 An Overview of HPF Fortran 90D High Performance Fortran (HPF) 4] is based on the research language Fortran 90D [5] and provides a minimal set of extensions to Fortran 90 to support the data parallel programming model 1 . Extensions incorporated into HPF Fortran 90D provide a means for explicit expression of parallelism and data mapping. 1 The data parallel programming model is defined as single threaded, ....

Geoffrey C. Fox, Seema Hiranandani, Ken Kennedy, Charles Koebel, Uli Kremer, Chau-Wen Tseng, and Min-You Wu, "Fortran D Language Specifications", Technical Report SCCS 42c, Northeast Parallel Architectures Center, Syracuse University, Syracuse NY 13244-4100, Dec. 1990, Available via WWW at http://www.npac.syr.edu.

No context found.

G. Fox, S. Hiranadani, K. Kennedy, C. Koelbel, U. Kremer, C. Tseng, and M. Wu. FortranD Language Specification. Technical Report SCCS-42c, Rice COMP TR90-141, 37p, 1991.

No context found.

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C. Tseng and M. Wu, `Fortran D language specification', Department of Computer Science Rice COMP TR90-141, Rice University (December 1990).

No context found.

Fox G., Hiranandani S., Kennedy K., Koelbel C., Kremer U., Tseng C. and Wu M., 1991. "Fortran D Language Specification", Rice U., TR90--141.

No context found.

G. C. Fox, et al., "Fortran D Language Specifications," Technical Report SCCS 42c, NPAC, Syracuse University, Syracuse, NY.

No context found.

G. Fox, S. Hiranandani, K. Kennedy, C. Koelbel, U. Kremer, C. Tseng, and M. Wu, `Fortran D language specification', Department of Computer Science Rice COMP TR90-141, Rice University, (December 1990).

First 50 documents  Next 50

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