D. B. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1993.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....view one big common file or many independent smaller files. Although several parallel programming models have been proposed in the past, of late, at least in the realm of massively parallel computing, there seems to be a convergence on the data parallel programming model [3] in the form of HPF [9] and Fortran 90 [4, 10] A natural way of executing a data parallel program is on a SIMD machine. However, a data parallel program may just as well be executed on an asynchronous MIMD machine, and there are several advantages to do so. The chief advantage is that it is possible to run several jobs ....

....data (PD) do not have to We assume, for simplicity, that a compiler or some preprocessing tool divides the public data in blocks in an optimized way, exploring data locality as much as possible. In fact, some tools have been recently proposed, such as Chaos [13] or the compiler directives of HPF [9] S virtual address physical address Figure 1. The OS virtual memory function . be necessarily divided among VPs. Take, for instance, the PRAM model that assumes a common memory equally accessible by all VPs. We will then redefine a parallel program as a collection of VPs (each composed of ....

D. B. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1(1):25 -- 42, 1993.

....One result of this is that a change in the data distribution method often requires a re write of the program. Writing programs at this level of detail also complicates performance analysis. An alternative approach is the use of a parallel language. One example is High Performance Fortran (HPF) [18]. The operation of the algorithm can be specified using a single thread of control, simplifying code development. Parallelism is derived from the partitioning of shared data sets across the processors (for example using block and cyclic distributions) Changing the data distribution is usually a ....

D. B. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1(1):17--42, February 1993.

....make FORTRAN 90 more convenient to use than the V RAM model; but, from a conceptual standpoint, they are superfluous. Linear arrays can be used to craft multi dimensional arrays, and all FORTRAN 90 intrinsics can be implemented on a V RAM indirectly in O(1) steps. High Performance Fortran (HPF) Lov93] is another data parallel programming language. It is merely an enhanced version of FORTRAN 90. The enhancements take the form of directives. The programmer uses these directives to suggest how the compiler should allocate each array among the various memory banks; the compiler tries, but is not ....

D.B. Loveman. High performance fortran. IEEE Parallel and Distributed Technology, 1(1):25 -- 42, 1993.

....high level languages (like C and Fortran) and try to recycle their already available sequential software. For these programmers, the extensions to existing languages or run time libraries are a viable alternative. 1.5. 3 High Performance Fortran The High Performance Fortran (HPF) initiative [20] seems to be a promising solution to solve the dusty deck problem of Fortran codes. However, it only supports applications that follow the SPMD paradigm and have a very regular structure. Other applications that are missing these characteristics and present a more asyn Section 1.5 Parallel ....

D. Loveman. High-Performance Fortran. IEEE Parallel and Distributed Technology, vol. 1 (1), February 1993.

.... of the neighbor s memory (main, cache or disk) Management policies must define: Page Placement (where to put it ) Page Identification (how to find it ) Page 1 It is assumed, for simplicity, that a compiler or some preprocessing tool,such as Chaos [P 95] the compiler directives of HPF [Lov93] or software caching [BL92] divides PD in blocks in an optimized way, exploring data locality as much as possible. Replacement; Write Strategies (write through, write back etc. Write Miss Strategies (to allocate or not to allocate ) These questions also apply to traditional VM, and some ....

D. B. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1(1):25 -- 42, 1993.

....of programming coarse grained parallel applications. These tools adopt a variety of approaches to add parallelism to existing languages (primarily C and Fortran) such as providing library calls (e.g. PVM [11] and Linda [6] loop parallelism (e.g. Myrias PAMS [14] extending a language (e.g. HPF [16]) and expressing the parallelism graphically (e.g. Hence [4] and Enterprise [19] More radical approaches include defining a new procedural language that supports parallelism (e.g. Orca [3] or alternative programming paradigms (such as a functional language like Sisal [9] However, given the ....

D. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, vol. 1, no. 1, pp. 25-42, 1993.

....processors. the performance of the program is limited by the fact that the workers share the same communication medium and by the low throughput and the high latency of the Ethernet. 5 Related Work Many parallel languages support only data parallelism, for example High Performance Fortran (HPF) [15], and pC [4] Likewise, many task parallel languages have been designed, such as Mentat [10] Compositional C [6] and Orca. A task parallel language can be used to write data parallel applications, but doing so is often awkward. For example, the implementation of Successive OverRelaxation in ....

D. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1(1), Feb. 1993.

....[105] provides data parallel extensions to C. It is based on the High Performance Fortran (HPF) standard, but goes beyond HPF in a number of areas with features such as dynamically distributed data structures and dynamically allocatable irregular arrays. 4.7. 4 HPF HPF (High Performance Fortran) [63, 70] was developed by the High Performance Fortran Forum, a coalition of academic and industrial groups. The basic idea behind HPF is to provide a set of extensions to Fortran 90, allowing data parallel programs to be written which achieve respectable performance on parallel machines. Several packages ....

D. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1(1):25--42, 1993.

....dynamic communications, just expressed somewhat differently. While the thesis explicitly ignores the issue of converting high level language inputs into structured communications, this process can be either helped or hindered by the choice of high level language. Some languages, such as HPF [46] or Crystal [15] provide high level communications operators; these operators easily allow for suitable communication schedules to be generated. Other languages, such as Mul T [38] and Linda [14] have less emphasis on high level operations and are harder to extract structured communications ....

David B. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1(1):25--42, February 1993.

....require the application to specify the manner in which different parts of the application are related in terms of dependencies or constraints. Examples of low level parallel programming systems are PVM [23, 12] MPI [13] and Linda [8] Examples of high level languages are PCN [9] CC [24] HPF [18], HeNCE [6] and CODE [21] Whereas low level systems give the application direct control over how parallelism is to be managed, they tend to be difficult to use and maintain. In some cases, they compromise the degree to which applications can be architecture independent. On the other hand, ....

....PCN uses explicit operators and shared variables, whereas GLU uses multidimensional data dependencies and functions. PCN requires use of recursion to express data parallelism, whereas GLU can express data parallelism using multidimensional data dependencies. HPF (High Performance Fortran) [18] is a parallel extension of Fortran 90 that supports SPMD programming. An HPF program is a Fortran 90 program with a single thread of control and a global address space. In addition, the programmer annotates the program with data mapping directives to achieve desired data locality and ....

David B. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1(1):25--42, February 1993.

....by students for their sequential programming (preferably C ) was available on the computers they use, was easy to learn by beginners, and was usable by students at all levels in many kinds of courses. Although many data parallel languages exist, including C , Fortran90, NESL [Ble93] and HPF [Lov93] they are difficult to use, are not similar to C , or are not easily portable to student computers. We found many research projects designing parallel C variants. C [LRV92] is perhaps the closest candidate, in that it supports a data parallel model, but it requires a new compiler and is not ....

David B. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1(1):25--42, February 1993.

....model and the physical machine. ffl The programming model Several parallel programming models have been proposed in the past. Of late, at least in the realm of massively parallel computing, there seems to be a convergence on the data parallel programming model [2, 7] in the form of HPF [12] and Fortran 90 [3, 16] Sabot [20] mentions that, in a survey of 120 parallel algorithms from three ACM Symposia on Theory of Computing (STOC) all were found to be data parallel. This preponderance of the data parallel programming model is perhaps because it allows one to express a large degree ....

David B. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1(1):25 -- 42, 1993.

....of a Restart object rereads a series of checkpointed Arrays into the compute nodes memories. In addition to supporting these various ArrayGroup objects, Panda allows different data distributions to be used in the compute node memories and I O node disks. For instance, using the HPF terminology [7], one could have a (BLOCK, distribution on the compute nodes, but for whatever reason, use a ( BLOCK) distribution when writing to disk. The purpose of this project was to retain most of Panda s high level interfaces with dealing with large arrays, but to change some of the actual ....

David B. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1(1):25-42, February 1993.

....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 ....

D. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1:25--42, February 1993.

....standardize on a single interface [26] It is widely believed that shared memory programming models are easier to use than message passing models. This belief is supported by the dominance of (small scale) shared16 memory multiprocessors in the market, and by the many efforts by compiler writers [32, 68, 56] and operating system developers [19, 62] to provide a shared memory programming model on top of message passing hardware. We focus on shared memory models in the remainder of this chapter. 0.3.2 Implementation of Shared Memory Models Any user level parallel programming model must deal with ....

....scientific computing, where multicomputers dominate the market, much attention has recently been devoted to the development of efficient shared memory programming models. Most notably, a large number of research groups have cooperated in the development of High Performance Fortran (HPF) [56]. HPF is intended not as an ideal language, but as a common starting point for future developments. To first approximation, it combines the syntax of Fortran 90 (including operations on whole arrays and slices) with the data distribution and alignment concepts developed in Fortran D [42] Users ....

D. B. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1(1):25--42, February 1993.

....would determine the best location for each datum at each point in time, and would place data accordingly. Compiler technology has not yet advanced to the point where this task is feasible; the current state of the art assumes a user specified distribution of data among processors (e.g. as in HPF [14]) Moreover, there will always be important programs for which reference patterns cannot be determined at compile time, e.g. because they depend on input data [16] Even for those applications in which compile time placement is feasible, it still seems possible that OS level placement will offer a ....

D. B. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1(1):25-- 42, February 1993.

....high level parallel programming languages. Parallel languages provide a common programming paradigm and memory access model across architectures by shifting (some of) the burden of architecture specific adaptations to the compiler. The most widely accepted example is High Performance Fortran (HPF) [12], which is now supported by a number of vendors. The initial version of the language was intended primarily for data parallel applications, but current developments are intended to broaden its applicability to irregular problems. Similar efforts involving parallel derivatives of C may be more ....

D. B. Loveman. High performance fortran. IEEE Parallel and Distributed Technology, 1(1), 2542, Feb. 1993.

....parallel programs. Finally, we introduce lock independent code motion, a new optimization technique that attempts to minimize the size of critical sections in the program. 1. Introduction Although recent advances in parallelizing compilers and data parallel languages have been impressive [4, 8], there are important problem domains for which parallelizing the best sequential algorithm or data layout yields sub optimal performance relative to an implementation that is explicitly parallel from the outset. Furthermore, popular systems like Java incorporate parallel constructs at the ....

D. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1(1):25--43, 1993.

.... memories according to programmer supplied directives, and to partition a computation among processors in a way that reduces communication [8] High Performance Fortran (HPF) for example, provides a programmer with template, align, and distribute directives to distribute arrays across processors [29]. Although a programmer usually partitions data, the compiler partitions the computation among processors. A common approach is owner computes, in which the processor holding a location (its home processor) computes and assigns the location s new values. At each assignment statement, processors ....

David B. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1(1):25--42, February 1993.

....by students for their sequential programming (preferably C ) was available on the computers they use, was easy to learn by beginners, and was usable by students at all levels in many kinds of courses. Although many data parallel languages exist, including C , Fortran90, NESL [Ble93] and HPF [Lov93] they are difficult to use, are not similar to C , or are not easily portable to student computers. We found many research projects designing parallel C variants. C [LRV92] is perhaps the closest candidate, in that it supports a data parallel model, but it requires a new compiler and is not ....

David B. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1(1):25--42, February 1993.

....advantages offered by its low level nature. 8.3 Higher level Interfaces There are numerous interfaces that are designed to allow programmers to describe their I O needs at a higher semantic level. These interfaces are sometimes tightly integrated into a particular language such as HPF [Lov93, BGMZ92, HPF93] or CMF [Thi94] There has also been great deal of research in designing libraries to support parallel I O. The focus of much of this research has been the support of distributed matrices, particularly for data parallel variants of Fortran [CBH 94, TBC 94, TG96] Other ....

David B. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1(1):25--42, February 1993.

....the hardware cache controllers are capable of filling misses from remote locations. Ideally, the compiler for a parallel language would determine the best location for each datum at each point in time. Most current compilers, however, employ a programmerspecified data distribution (e.g. as in HPF [12]) Moreover, there will always be important programs for which reference patterns cannot be determined at compile time, e.g. because they depend on input data [13] Even when compile time placement is feasible, it still seems possible that OS level placement will offer a simpler, acceptable ....

D. B. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1(1):25--42, Feb. 1993.

....approach as followed by MPI and OpenMP is to provide a small set of modifications and or extensions to an existing high level programming language such as C C , Fortran, or Java. Probably the most popular example of a language that has adopted this approach is HPF (High Performance Fortran [29]) Also, many alternative extensions and modifications to C exist [30] of which Compositional C [31] and Mentat [32] are the most significant examples. Irrespective of language design and compilation issues, for users of such languages the most important problem is that it is often required ....

Loveman D. High Performance Fortran. IEEE Parallel and Distributed Technology 1993; 1:25--42.

No context found.

D. B. Loveman. High Performance Fortran. IEEE Parallel and Distributed Technology, 1993.

No context found.

D. Loveman. High Performance Fortran, IEEE Parallel and Distributed Technology, vol. 1, no. 1, pp. 25-42, 1993. 20

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