High Performance FORTRAN Language Specification. Technical Report, Version 2.0.ffi, Rice University, Houston, TX, October 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

.... overhead [9] All of these problems can be formulated as queries based on a set of linear and non linear constraints I defined over loop variables and parameters (loop invariants) which are commonly derived from loop bounds, conditional statements, data and work distribution specifications [16, 2], etc. For instance, I is given by f1 J 1 N , N= 2 J 1 ) J 2 N , J 1 J 2 Ng for Example 1.2. The number of solutions of I can then be used to answer all problems mentioned above. Furthermore, many analyses based on systems of constraints I are significantly alleviated if I is simplified by ....

....in its local memory. The work contained in the program is distributed according to a data distribution, and a processor performs computations that define the data elements that the processor owns. Both data and work distribution are commonly specified through directives, such as HPF directives [16]. If a processor requires data that is owned by a remote processor, the non local data is accessed through interprocessor communication which is automatically optimized and inserted by VFCS. ffl Let I be a set of non linear and linear constraints (equalities and inequalities) defined over loop ....

High Performance FORTRAN Language Specification. Technical Report, Version 1.0, Rice University, Houston, TX, May 1993.

....results. Finally, some concluding remarks are made and future work is outlined. 2 Compilation and Programming Model The programs which are estimated by P T are based on the underlying compilation and programming model of VFCS ( 2] which is similar to most High Performance Fortran (HPF [8]) style compilers such as Fortran D ( 9] The parallelization strategy of VFCS is based on data decomposition in conjunction with the Single Program Multiple Data (SPMD) programming model. With this method, the data arrays in the original program are each partitioned and mapped to the processors ....

....implemented by VFCS through message passing. The input to VFCS are Vienna Fortran programs ( 2] Vienna Fortran is a machine independent language extension to Fortran77, which provides annotations for the specification of data distributions. VFCS supports a limited frontend for HPF programs ([8]) which includes PROCESSORS, DISTRIBUTE, ALIGN, and INHERIT HPF directives. Both Vienna Fortran and HPF programs are translated to the same internal program representation. This allows various compiler transformations and integrated tools which also accounts for P T to be applied to the ....

High Performance FORTRAN Language Specification. Technical Report, Version 2.0.ffi, Rice University, Houston, TX, October 1996.

....to exploit this power comprises a collection of concepts that are briefly described in the following sections (cf. also Fig. 1) The program to be executed is parallelized using proven program parallelization mechanisms designed for data mapping, as defined e.g. in High Performance Fortran(HPF) [1]) The processors involved can be located anywhere on a global network; they are used to store local portions of the mapped data, i.e. have the same function as processors declared in an HPF program. Unmapped data, which in HPF implementations is broadcasted over all cooperating processors, is ....

....input is: reg. sect. alignm. layout DeltaM ffi receiver: 1 : 22 : 3) 1; 0) 6; 3) f3; 3g f1; 1; 1g f0; 0; 0g sender: 0 : 7 : 1) 1; 0) 8; 1) f1; 1; 1; 1; 1; 1; 1; 1g f0g f0g The result produced is: recv B(1:22:3) 0] 0] f1; 4; 7; 10g sendC [0] 0] f0; 1; 6; 7g recv B(1:22:3) [1][0] f1; 4g sendC [0] 1] f2; 3g recv B(1:22:3) 2] 0] f1; 4g sendC [0] 2] f4; 5g 3.4 Creation of Execution Blocks In order to minimize the number of calls by which the conductor initializes the execution of operations on cooperating servant collections, the compiler ties together ....

[Article contains additional citation context not shown here]

High Performance Fortran Forum: High Performance Fortran Language Specification. Technical Report, Version 2.0.delta, Rice University, October 20, 1996.

....transformations and optimizations applied by a compiler. In this paper we introduce P 3 T , the successor tool of P 3 T [22, 15, 16] which models programs, code transformations, and parallel and distributed architectures. The input programs of P 3 T are written in High Performance Fortran [27, 1] which represents the de facto standard of high level data parallel programming. Moreover, P 3 T analyzes Fortran90 message passing programs generated by the underlying compiler (VFC [2] which can be executed on parallel and distributed machines such as network of workstations. P 3 T models ....

....refers to the sequential computation time overhead. In the worst case predicted computation times are off the measured values by 10 . Figure 7 shows estimated and measured work distribution values for two different parallel versions of the Cholesky factorization (BLOCK and CYCLIC distribution [27] of the first dimension of array A) that has been executed on 16 processors. For BLOCK distribution, the predicted work distribution values are off the measured results in the worst case by 0.6 . The Cholesky factorization based on CYCLIC distribution can yield estimation errors of up to 35 for ....

[Article contains additional citation context not shown here]

High Performance FORTRAN Language Specification. Technical Report, Version 2.0.ffi, Rice University, Houston, TX, October 1996.

....applied by a compiler. P 3 T , the successor tool of P 3 T [5, 6] is a performance estimator for distributed and parallel programs which models programs, code transformations, and parallel and distributed architectures. The input programs of P 3 T are written in High Performance Fortran [10] which represents the de facto standard of high level data parallel programming. Moreover, P 3 T analyzes Fortran90 message passing programs generated by the underlying compiler (VFC [1] to be executed on parallel and distributed architectures (e.g. network of workstations) P 3 T ....

High Performance FORTRAN Language Specification. Technical Report, Version 2.0.ffi, Rice University, Houston, TX, October 1996.

....transformations and optimizations applied by a compiler. In this paper we introduce P 3 T , the successor tool of P 3 T [20, 13, 14] which models programs, code transformations, and parallel and distributed architectures. The input programs of P 3 T are written in High Performance Fortran [25, 1] which represents the de facto standard of high level data parallel programming. Moreover, P 3 T analyzes Fortran90 message passing programs generated by the underlying compiler (VFC [2] which can be executed on parallel and distributed machines such as network of workstations. P 3 T models ....

High Performance FORTRAN Language Specification. Technical Report, Version 2.0.ffi, Rice University, Houston, TX, October 1996.

....that current parallelizing compilers are not very effective in optimizing parallel programs that fully utilize the target multiprocessor system. The poor performance of many compiler analyses can be attributed to ineffective parallelization of programs (for instance, High Performance Fortran HPF [34] and Fortran 90 [40] that have a strong potential for unknowns such as number of processors and sizes of 1 allocateable arrays. Some compilers have a problem to make the relationship between a program s problem size (input data, array and machine sizes, etc. and analysis information explicit. ....

.... of our symbolic evaluation framework which is used as part of the Vienna Fortran Compilation System (VFCS) 5] a parallelizing compiler for distributed memory architectures and P 3 T [21, 22] a performance estimator to parallelize and optimize High Performance Fortran programs [34, 5] for distributed memory architectures. The organization of this paper is as follows. Preliminaries are presented in Section 2. In Section 3, we describe our symbolic evaluation framework. This includes techniques for computing path conditions and values of symbolic expressions. We will further ....

[Article contains additional citation context not shown here]

High Performance FORTRAN Language Specification. Technical Report, Version 2.0.ffi, Rice University, Houston, TX, October 1996.

....by a compiler. P 3 T , the successor tool of P 3 T [15, 11, 12] is a performance estimator for distributed and parallel programs which models programs, code transformations, and parallel and distributed architectures. The input programs of P 3 T are written in High Performance Fortran [17, 1] which represents the de facto standard of high level data parallel programming. Moreover, P 3 T analyzes Fortran90 message passing programs generated by the underlying compiler (VFC [2] and which can execute on parallel and distributed machines such as network of workstations. P 3 T ....

....refers to the sequential computation time overhead. In the worst case predicted computation times are off the measured values by 10 . Figure 3 shows estimated and measured work distribution values for two different parallel versions of the Cholesky factorization (BLOCK and CYCLIC distribution [17] of the first dimension of array A) that has been executed on 16 processors. For BLOCK distribution, the predicted work distribution values are off the measured results in the worst case by 0.6 . The Cholesky factorization based on CYCLIC distribution can yield estimation errors of up to 35 for ....

[Article contains additional citation context not shown here]

High Performance FORTRAN Language Specification. Technical Report, Version 2.0.ffi, Rice University, Houston, TX, October 1996.

....programs for distributed and parallel systems is an error prone and timeconsuming process that may involve many cycles of code editing, compiling, executing, and performance analysing. Many different programming paradigms such as explicit message passing [15] High Performance Fortran (HPF) [20], OpenMP [9] Java RMI [39] and HPC [26] have been introduced for distributed and parallel architectures. A trade off is implied by the programming paradigm employed. On the one hand, programming at a low level (i.e. message passing paradigm) enables the programmer to fully exploit and control ....

....= H(I,1:I) C1I(1,1:I) C2I(1,I) 260 CONTINUE . Figure 13. HNS based on HPF Fortran90 array operations 5.1 Parallelization We used VFC [2] to generate a code in two different parallel HNS versions which are based on HPF. In both versions H, the main HNS array, has been distributed CYCLIC [20] in the second dimension onto the maximum number of processors (HPF intrinsic function NUMBER OF PROCESSORS) that are available on a given architecture. In order to achieve good work distribution, CYLCIC distribution has been chosen according to triangular loop iteration spaces. In the first ....

High Performance FORTRAN Language Specification. Technical Report, Version 2.0.ffi, Rice University, Houston, TX, October 1996.

....we use P 3 T [7, 8, 6, 9] an accurate and effective performance estimation tool for distributed memory parallel programs. P 3 T is a static performance estimator that analytically estimates the performance of data parallel programs (subset of Vienna Fortran [31] High Performance Fortran [20], Fortran90 and Fortran77) at compile time without using simulation. P 3 T has been developed to guide the selection of efficient data distribution strategies and profitable code transformations. A variety of perfor3 mance parameters each reflecting a different performance aspect are estimated: ....

High Performance FORTRAN Language Specification. Technical Report, Version 2.0.ffi, Rice University, Houston, TX, October 1996.

....be formulated as queries based on a set of linear and non linear constraints I. The constraints of I are defined over loop variables and parameters (loop invariants) which are commonly derived from loop bounds, conditional statements, data declarations, data and work distribution specifications [17, 1], etc. For instance, I is given by f1 J 1 N , N= 2 J 1 ) J 2 N , J 1 J 2 Ng for example 2. The number of solutions of I can then be used as a basis for the solution of all problems mentioned above. Furthermore, many analyses based on systems of constraints I are significantly alleviated if ....

....program is distributed according 4 THOMAS FAHRINGER to a data distribution, and a processor performs computations that define the data elements that the processor owns. Both data and work distribution are commonly specified through directives, such as HPF (High Performance Fortran) directives [17]. If a processor requires data that is owned by a remote processor, the non local data is accessed through interprocessor communication which is automatically optimized and inserted by VFCS. ffl Let I be a set of non linear and linear constraints (equalities and inequalities) defined over loop ....

High Performance FORTRAN Language Specification. Technical Report, Version 2.0.ffi, Rice University, Houston, TX, October 1996.

....a path from e to x. An instruction may write or use data references (array or scalar variables) A program is executed by a set of processors P . The computations and data of a program are distributed to one or more processors in P according to the underlying data and work distribution strategy [3, 14, 15]. Every processor contains a private copy of all scalars. Each data element is owned by one or more processors. Our strategy covers both non local uses and non local writes. Non local data must be fetched before it is used by a processor. If a processor writes data that is owned by another ....

High Performance FORTRAN Language Specification. Technical Report, Version 1.0, Rice University, Houston, TX, May 1993.

....on experimental results. Finally, some concluding remarks are made and future work is outlined. 2 Compilation and Programming Model The programs which are estimated by P 3 T are based on the underlying compilation and programming model of VFCS ( 2] which is a High Performance Fortran (HPF [8]) style compiler. The parallelization strategy of VFCS is based on data decomposition in conjunction with the Single Program Multiple Data (SPMD) programming model. With this method, the data arrays in the original program are each partitioned and mapped to the processors of the target ....

....by VFCS through message passing. The input to VFCS are Vienna Fortran programs ( 2] Vienna Fortran is a machine independent language extension to Fortran77, which provides annotations for the specification of data distributions. Furthermore, VFCS supports a limited frontend for HPF programs ([8]) which includes PROCESSORS, DISTRIBUTE, ALIGN, and INHERIT HPF directives. Both Vienna Fortran and HPF programs are translated to the same internal program representation. As most compiler analysis which also accounts for P 3 T is based on this internal representation, it can be applied ....

High Performance FORTRAN Language Specification. Technical Report, Version 1.0, Rice University, Houston, TX, May 1993.

....on defining stereotypical patterns of parallel computation which can be put together. We briefly look at each of these in turn. In the first category lie systems which have extended a procedural language with explicit constructs for parallelism such as C [8] and High Performance Fortran [10], which is the latest language that combines Fortran 90 with enhanced user defined data distribution. A similar idea is the provision of standard message passing library extensions, such as MPI and PVM, which enable much flexibility through facilitating the use of existing sequential languages ....

High performance fortran language specification, version 1.0. Technical Report, May 1993. Rice University

....approach in which the user and the compiler share responsibility for exploiting parallelism. The main goal of the group has been to design a high level set of standard extensions to Fortran called, High Performance Fortran (HPF) intended to exploit a wide variety of parallel architectures [28, 40]. The HPF extensions allow the user to carefully control the distribution of data across the memories of the target machine. However, the computation code is written using a global name space with no explicit message passing statements. It is then the compiler s responsibility to analyze the ....

....prefix, HPF , needs to be removed to retain a correct HPF program. HPF also introduces some new language syntax in the form of data parallel execution statements and a few new intrinsics. z This chapter is partially based on the High Performance Fortran Language Specification draft document [28] which has been jointly written by several of the participants of the High Performance Fortran Forum. Also, the specification (as described here) are still under review and may change when the final document is released. Features of High Performance Fortran In this subsection we provide a brief ....

[Article contains additional citation context not shown here]

High Performance FORTRAN Language Specification. Technical report, Rice University, 1993.

....programs. It cannot be employed for irregular problems, which require runtime analysis. Shifting performance estimation into runtime to support runtime optimization will be addressed in future work. P 3 T is currently being extended to estimate the performance of High Performance Fortran (HPF [21]) programs using a separate frontend under the VFCS. Ongoing work to fine tune the estimator for a larger set of optimizing transformations and to evaluate it for several other distributed memory architectures will further enhance the usefulness of P 3 T . Additional information on our ....

High Performance FORTRAN Language Specification. Technical Report, Version 1.0, Rice University, Houston, TX, May 1993.

No context found.

Performance Comparison of iSCSI and NFS IP Storage Protocols. Technical report, TechnoMages, Inc.

....users to design programs for massively parallel computers much as they are accustomed to on a sequential machine, have been the subject of intense discussion and research activity in recent months. Vienna Fortran [3, 4] is one of several proposals put forth for such a set of language extensions [5, 6, 9, 12, 13]. A number of features of Vienna Fortran have since been adopted by the High Performance Fortran Forum. One of these is the concept of static and dynamic distributions of arrays in a program, although the details of these features are not the same in High Performance Fortran (HPF) The language ....

....of arrays, and does not include any additional constructs which might enable the user to control or structure the use of dynamic distributions. As far as we are aware, the Fortran D implementation does not yet provide for dynamic data distributions. The High Performance Fortran proposal [9] includes static and dynamic distributions in much the same way that Vienna Fortran does and has included a small set of distribution queries in the language constructs. It has REALIGN and REDISTRIBUTE directives to permit independent redistribution and realignment of arrays during execution. ....

High Performance FORTRAN Language Specification. Technical report, Rice University, May 1993.

No context found.

High Performance FORTRAN Language Specification. Technical Report, Version 2.0.ffi, Rice University, Houston, TX, October 1996.

No context found.

High Performance FORTRAN Language Specification. Technical Report, Version 2.0.ffi, Rice University, Houston, TX, October 1996.

No context found.

High Performance FORTRAN Language Specification. Technical Report, Version 2.0.ffi, Rice University, Houston, TX, October 1996.

No context found.

High Performance Fortran Language Speci cation. Technical report, Rice University, Texas, USA, 1993.

No context found.

High Performance FORTRAN Language Specification. Technical Report, Version 1.0, Rice University, Houston, TX, May 1993.

No context found.

High Performance Fortran Language Specification, Technical Report, Rice University, Texas, USA, 1993.

No context found.

Performance Fortran Language Speci#cation. Technical report, Rice University, November 1994.

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