I. Foster and K.M. Chandy. Fortran M: A Language for Modular Parallel Programming. Journal of Parallel and Distributed Computing, 25(1):24--35, April 1995.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....into appropriate tasks such that the target platform is exploited efficiently. Specifically, a large number of processors requires a large number of tasks which is often not available within an application. Task parallelism is provided by task parallel languages such as Linda [45] Fortran M [75], Fx [96] CC [52] and Orca [11] With the growing popularity of distributed object computing a number of middleware [28] systems that provide task parallel features have been developed. Examples include CORBA [155] and Microsoft s DCOM [110] Traditionally, either data parallelism or task ....

....Moreover, the compiler approach allows an easier integration of new language features, such as those described in Chapter 3, whose implementation and integration in the Opus system is discussed below. Related coordination approaches that make use of a compiler are, for example, Fx [96] Fortran M [75], or Orca [11] Fx does not need any runtime support since it has a quite simple tasking model similar to the tasking facilities in HPF 2 (in fact, HPF 2 took over many ideas introduced in Fx) Fortran M is based upon Nexus which fits perfectly into the Fortran M model. However, Fortran M is a ....

I. Foster and M. Chandy. Fortran M: A Language for Modular Parallel Programming. Journal of Parallel and Distributed Computing, 26(1):24--35, 1995.

....models support the programmer in writing efficient programs without dealing too much with the underlying communication and coordination details of a specific parallel machine. Language approaches include Braid, Fortran M, Fx, Opus, and Orca, see [2] for an overview and comparison. Fortran M [8, 7] allows the creation of processes which can communicate with each other by predefined channels and which can be combined with HPF Fortran for a mixed task and data parallel execution. The organization of communication in Fortran M is different than within the Tlib library, since communication in ....

I. Foster and K.M. Chandy. Fortran M: A Language for Modular Parallel Programming. Journal of Parallel and Distributed Computing, 25(1):24--35, April 1995.

....mixed task and data parallelism has emerged as a natural solution. We show that applying both data and task parallelism can improve the speedup at the application level. There have been considerable effort in adding taskparallel support to data parallel languages, as in Fx [15] Fortran M [4] or Paradigm HPF [12] or (b) adding data parallel support to task parallel languages such as in Orca [7] In order to fully exploit the potential advantage of the mixed task and data parallelism, efficient support for task and data parallelism is a critical issue. This can be done not only at the ....

T. Foster and K. Chandy. Fortran m: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 26:24--35, 1995.

....matrix multiplication) running on a cluster of workstations. 1 Introduction Recently, it has been shown that exploiting both task and data parallelism in a program to solve very large computational problems yields better speedups compared to either pure task parallelism or pure data parallelism [1, 5, 14, 22]. The main reasons is that both task and data parallelism are relatively limited, and therefore using only one of them needlessly bounds the achievable performance. Thus, exploiting mixed task and data parallelism has emerged as a natural solution. There have been considerable effort in (a) ....

....limited, and therefore using only one of them needlessly bounds the achievable performance. Thus, exploiting mixed task and data parallelism has emerged as a natural solution. There have been considerable effort in (a) adding task support to data parallel languages, as in Fx [20, 22] Fortran M [5] or Paradigm HPF [14] or (b) adding data parallel support to task parallel languages such as in Orca [1] In order to fully exploit the potential advantage of mixed task and data parallelism, efficient support for mixed task and data parallel scheduling is a critical issue. Mixed This ....

T. I. Foster and K. M. Chandy. Fortran M: A language for modular parallel programming. J. Parallel and Distributed Computing, 26:24--35, 1995.

....limited, and therefore using only one of them bounds the achievable performance. In order to fully exploit the potential advantage of the mixed task and data parallelism, efficient support for task and data parallelism is a critical issue. This can be done not only at the compiler level [5, 7, 8, 9], but also at the application level and applications from the image processing field are very suitable for this technique. Mixed task and data parallel techniques use a directed acyclic graph, in the literature also called a Macro Dataflow Graph (MDG) 5] in which data parallel tasks are the ....

T.I. Foster, and K.M. Chandy: Fortran M: A language for modular parallel programming, in Journal of Parallel and Distributed Computing, 26:24-35, 1995.

....relatively limited, and therefore using only one of them bounds the achievable performance. In order to fully exploit the potential advantage of the mixed task and data parallelism, ecient support for task and data parallelism is a critical issue. This can be done not only at the compiler level [5, 7, 8, 9], but also at the application level and applications from the image processing eld are very suitable for this technique. Mixed task and data parallel techniques use a directed acyclic graph, in the literature also called a Macro Data ow Graph (MDG) 5] in which data parallel tasks are the nodes ....

T.I. Foster, and K.M. Chandy: Fortran M: A language for modular parallel programming, in Journal of Parallel and Distributed Computing, 26:24-35, 1995.

....mixed task and data parallelism has emerged as a natural solution. We show that applying both data and task parallelism can improve the speedup at the application level. There have been considerable e ort in adding task parallel support to data parallel languages, as in Fx [10] Fortran M [11] or Paradigm HPF [7] or adding data parallel support to task parallel languages such as in Orca [12] In order to fully exploit the potential advantage of the mixed task and data parallelism, ecient support for task and data parallelism is a critical issue. This can be done not only at the ....

T.I. Foster, and K.M. Chandy: Fortran M: A language for modular parallel programming, in Journal of Parallel and Distributed Computing, 26:24-35, 1995.

.... and algorithmic skeletons for which parallelism is implicitly available as independent expressions can be evaluated in parallel, and data parallel languages like HPF (High Performance Fortran) 5] or NESL [3] Other language approaches include Braid, Fortran M, Fx, Opus, and Orca [1] Fortran M [7] allows the creation of processes which can communicate with each other by predefined channels and which can be combined with HPF Fortran for a mixed task and data parallel execution. The Fx model expresses task parallelism by providing declaration directives to partition processors into ....

I. Foster and K.M. Chandy. Fortran M: A Language for Modular Parallel Programming. Journal of Parallel and Distributed Computing, 25(1):24--35, April 1995.

....Research (NWO) under grant number SION 2519 61233 005. 1 1 Introduction Recent studies has shown that, for a large class of large computational applications, exploiting both task and data parallelism yields better speedups compared to either pure task parallelism or pure data parallelism [2, 7, 21, 18, 27]. In large applications, pure data parallelism is relatively limited, especially for irregular applications. Task parallelism, on the other hand, can indeed expose a degree of parallelism that is comparable to the combined task and data parallelism. However, this can only be achieved using a huge ....

....and data parallelism within an application involves (a) good support at the language level for both task and data parallelism, and (b) a good scheduler. At the language level, there has been a considerable effort in (a) adding task support to data parallel languages, as in Fx [25, 27] Fortran M [7] or Paradigm HPF [18] or (b) adding data parallel support to task parallel languages such as in Orca [2] Efficient support for mixed task and data parallel scheduling is a critical issue for fully exploiting the potential advantage of the mixed task and data parallelism. Mixed task and data ....

T. I. Foster and K. M. Chandy. Fortran M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 26:24--35, 1995.

....matrix multiplication) running on a cluster of workstations. 1 Introduction Recently, it has been shown that exploiting both task and data parallelism in a program to solve very large computational problems yields better speedups compared to either pure task parallelism or pure data parallelism [1, 5, 14, 22]. The main reasons is that both task and data parallelism are relatively limited, and therefore using only one of them needlessly bounds the achievable performance. Thus, exploiting mixed task and data parallelism has emerged as a natural solution. There have been considerable effort in (a) ....

....limited, and therefore using only one of them needlessly bounds the achievable performance. Thus, exploiting mixed task and data parallelism has emerged as a natural solution. There have been considerable effort in (a) adding task support to data parallel languages, as in Fx [20, 22] Fortran M [5] or Paradigm HPF [14] or (b) adding data parallel support to task parallel languages such as in Orca [1] In order to fully exploit the potential advantage of mixed task and data parallelism, efficient support for mixed task and data parallel scheduling is a critical issue. Mixed This ....

T. I. Foster and K. M. Chandy. Fortran M: A language for modular parallel programming. J. Parallel and Distributed Computing, 26:24--35, 1995.

....mixed task and data parallelism has emerged as a natural solution. We show that applying both data and task parallelism can improve the speedup at the application level. There have been considerable effort in adding taskparallel support to data parallel languages, as in Fx [15] Fortran M [4] or Paradigm HPF [12] or (b) adding data parallel support to task parallel languages such as in Orca [7] In order to fully exploit the potential advantage of the mixed task and data parallelism, efficient support for task and data parallelism is a critical issue. This can be done not only at the ....

T. Foster and K. Chandy. Fortran m: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 26:24--35, 1995.

....we could run a preprocessor over Java, but this is not a natural procedure in existing Java environments 5 Scienti c programmers have expected to program inter process communication at a higher level, if at all. More successful schemes include Message passing through language level support [24, 15] or higher level library interfaces [14] Data parallelism. We restrict the de nition of data parallelism to cover algorithms that achieve parallelism through operations on distributed arrays. Synchronization is usually limited to bulk synchronization occuring naturally through collective ....

I. Foster and K. M. Chandy. Fortran M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 26(1):24, 1995.

....in scheduling communications that is missing from raw C. We will give an example of socket programming in section 3. 1, but traditionally this has not been a popular paradigm in the parallel processing world, where more succesful schemes include Message passing through language level support [21, 13] or higher level library interfaces [12] Data parallelism, which we take to mean the style of programming in which parallelism is achieved through operations on distributed arrays, with synchronization typically limited to bulk synchronization occuring naturally through collective array ....

I. Foster and K. M. Chandy. Fortran M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 26(1):24, 1995.

....We believe that our work is complementary to that of Anderson et al. Flick can leverage the benefits of an efficient transport system to produce optimized communication stubs. For parallel applications, there are a large number of specialized programming languages such as CC [6] Fortran M [15], and Split C [8] In most of these cases the language handles marshaling and unmarshaling of parameters. However, it is our belief that the techniques used by Flick, and possibly even its code, could be incorporated into the compilers for these languages to substantially reduce presentation layer ....

I. T. Foster and K. M. Chandy. Fortran M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 25(1), Feb. 1995.

....matrix multiplication) running on a cluster of workstations. 1 Introduction Recently, it has been shown that exploiting both task and data parallelism in a program to solve very large computational problems yields better speedups compared to either pure task parallelism or pure data parallelism [2, 7, 18, 27]. The main reason is that both task and data parallelism are relatively limited, and therefore using only one of them needlessly bounds the achievable performance. Thus, exploiting mixed task and data parallelism has emerged as a natural solution. There have been considerable effort in (a) adding ....

....limited, and therefore using only one of them needlessly bounds the achievable performance. Thus, exploiting mixed task and data parallelism has emerged as a natural solution. There have been considerable effort in (a) adding task support to data parallel languages, as in Fx [25, 27] Fortran M [7] or Paradigm HPF [18] or (b) adding data parallel support to task parallel languages such as in Orca [2] This research is partially granted by the Netherlands Computer Science Foundation (SION) with financial support from the Netherlands Organization for Scientific Research (NWO) under grant ....

T. I. Foster and K. M. Chandy. Fortran M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 26:24--35, 1995.

.... combinations of two di erent archetypes (the mesh and mesh spectral) two di erent architectures (an IBM SP2 at Argonne National Labs using a straight interconnect, and a network of 166 MHz Pentium personal computers connected by 100Mbps ethernet) and two di erent languages libraries (Fortran M [17] and Fortran with MPI [18] Environment. Experiments were performed on otherwise unloaded computer nodes, but in an environment in which communications hardware was also supporting other users. Since execution times were consistent across multiple runs, we assume that this sharing of ....

I. T. Foster and K. M. Chandy. FORTRAN M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 26(1):24-35, 1995.

....of data objects distributed across processors. The task parallel programming model takes a different approach in which programs consist of a number of asynchronous, independent, communicating parallel processes. Task parallel languages such as CC [41] CHARM [125] CHARM [90] Fortran M [74], and Linda [39] define a set of mechanisms that coordinate process execution and communication among autonomous tasks. Task parallelism provides no explicit support for data decomposition. Task parallelism is ideally suited for computations integrating various heterogenous operations, such as a ....

....adaptive mesh methods. Of course, there are many applications that LPARX does not address. LPARX does not support the unstructured methods targeted by CHAOS [60] nor does it provide the dynamic irregular data types of Multipol [40] It cannot handle the task parallelism of CC [41] or Fortran M [74]. LPARX applies only to problems with irregular, block structured data exhibiting coarse grain data parallelism. Recent work with the Structural Abstraction (SA) model [13] extends the LPARX ideas to address other classes of irregular scientific applications (e.g. unstructured methods) Another ....

[Article contains additional citation context not shown here]

I. Foster and K. M. Chandy, Fortran M: A language for modular parallel programming, Journal of Parallel and Distributed Computing, (to appear).

.... whole) Therefore, a parallel program consists of two parts: a computation model and a coordination model [72] In his thesis, Yang [171] has suggested that coordination can be achieved using compiler directives as in Vienna Fortran [12] shared data pool as in Linda [31] channels as in Fortran M [69] and skeletons as in P 3 L [62, 125] There is plenty of work on coordination models, languages, theory, architecture and applications and it is beyond the scope of this thesis. In general, this 11 An example from real life is coordination between airplanes. Here airplanes are entities, ether ....

I. Foster and K. M. Chandy. Fortran M: A Language for Modular Parallel Programming. Journal of Parallel and Distributed Computing, 25(1), 1995.

....archetype, we provide a library of high level routines that encapsulate the details of parallelism. This library facilitates more rapid development of applications. The current implementation is based on Fortran 77 (including a few of the most widely supported extensions) and uses either Fortran M [4, 1] or MPI [6] for parallelism. Fortran M may be obtained via anonymous FTP from Argonne National Labs; many implementations of MPI are available. To use the archetype implementation, the application developer must provide: ffl A main program, typically consisting primarily of a call to spawn. ffl ....

.... 6 Implementation specifics As noted in x2, the current implementation is based on Fortran 77 (including a few of the most widely supported extensions long variable names, lower92 case source, and the do enddo construction) and uses either Fortran M ([4, 1], available via anonymous FTP from Argonne National Labs) or MPI ( 6] widely available) for parallelism. 6.1 Compiling and linking programs 6.1.1 Fortran M version To compile and link a program based on the Fortran M implementation, assuming the Fortran M compiler fm is installed: ffl Compile ....

I. T. Foster and K. M. Chandy. FORTRAN M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 26(1):24--35, 1995.

.... (mesh and mesh spectral) two different architectures (an IBM SP2 at Argonne National Labs using a straight interconnect, and a network of 166 MHz Pentium personal computers connected by 100Mbps Ethernet) and two different languages libraries (Fortran M and Fortran with MPI) Fortran M [FC95] consists of a small set of extensions to Fortran for modular parallel programming. In Fortran M, tasks and channels are represented explicitly, allowing the design of structured, unstructured, and asynchronous communication patterns for task parallel programs. In addition, Fortran M gives ....

I. T. Foster and K. M. Chandy. FORTRAN M: A Language for Modular Parallel Programming. Journal of Parallel and Distributed Computing, 26(1):24--35, 1995.

....the same (parameterized) code in a loosely synchronous fashion. Synchronization and exchange of data is handled by the compiler by inserting appropriate code. Languages providing support for managing task parallelism at a high level include LINDA [3] Strand [29] CC [11] and Fortran M [28]. Recently, a major research effort has been put on the issue of integrating task and data parallelism. Examples include Orca [7] Fx [76, 77] Approved Extensions for Data and Task Parallelism in HPF 2.0 [25] and Opus [12] In this work we develop compiler optimizations for reducing remappings ....

I. Foster and K.M. Chandy. Fortran M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 26, 1995.

....vector operators in Fortran 90 [5] or the intrinsic functions in various parallel Fortran systems. These constructs can be regarded as a set of parallel libraries encapsulating special forms of parallel computation. Co ordination Constructs: The examples include the port declarations in Fortran M[11], data distribution directives in Fortran D[9, 10] Vienna Fortran[16] and HPF[6] Rather than specifying computational tasks, these constructs are used to organise Fortran program components into a parallel ensembles. Co ordination constructs specify the organisation of a parallel program. ....

....parallel tasks is obviously necessary. This leads to the most recent efforts on developing various parallel Fortran system for the shared virtual memory computation model by using compiler directives to co ordinate parallel tasks. 2. 4 Co ordination By Channels DP77 [14] and Fortran M [11] are two parallel Fortran languages exploiting task parallelism. These languages comprise a small set of message passing extensions to Fortran and provides a message passing task parallel programming model. In Fortran M, concurrently executable processes are created using the PROCESSES or ....

Ian Foster, K. M. Chandy. Fortran M: A Language for Modular Parallel Programming. Technical Report, Argonne National Laboratory, 1992

....further models. Related work on performance prediction includes the studies in the context of parallelizing compilers [9, 21] as well as other prediction models [7, 5] Several research groups working on parallelizing compilers have included support to combine task and data parallelism. Fortran M [16, 11] allows the creation of processes which can communicate with each other by predefined channels and which can be combined with HPF Fortran for a mixed task and data parallel execution. In contrast to the Fortran M approach, the TwoL approach is more directed towards the combination of fine and ....

I. Foster and K.M. Chandy. Fortran M: A Language for Modular Parallel Programming. Journal of Parallel and Distributed Computing, 1995.

....CSP (Communicating Sequential Processes) 13, 14] initiated the area of distributed memory concurrent programming, which is now a large area of research. Many of the ideas have been incorporated into a variety of concurrentprogramming languages, including Concurrent ML [17] Facile [12] Fortran M [9], and occam [20] Common to these languages are two main concepts, synchronous communication over channels and non deterministic choice, that provide a powerful message passing abstraction. Briefly, message passing is provided by basic send and receive primitives. They are synchronous in that the ....

....implementation of general non deterministic communication over synchronous channels. This problem is typically ignored by either disallowing distributed systems (e.g. Concurrent ML [17] and Facile [12] or restricting the nondeterministic choice construct to allow only receives (e.g. Fortran M [9] and occam [20] We clearly do not want to restrict ourselves to sequential systems, since concurrent programs often have a large amount of parallelism that should be exploited using a distributed system. The so called generalized alternative construct, that is, one that supports both send and ....

I. Foster and K. M. Chandy. Fortran M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 26(1):21--35, 1995.

....compilers [7, 12] as well as other prediction models [6] In the following, we compare the TwoL approach to similar approaches in the area of parallelizing compilers. Several research groups working on parallelizing compilers have included support to combine task and data parallelism. Fortran M [9, 8] allows the creation of processes which can communicate with each other by predefined channels and which can be combined with HPF Fortran for a mixed task and data parallel execution. In contrast to the Fortran M approach, the TwoL approach is more directed towards the combination of fine and ....

I. Foster and K.M. Chandy. Fortran M: A Language for Modular Parallel Programming. Journal of Parallel and Distributed Computing, 1995.

....lead to an a priori estimation of the prospective gain of a parallel implementation. Thus, the programmer can decide on the benefits of a parallel implementation before actually performing it. An important application area for analytical performance prediction mechanisms are compilers for DMMs [21, 4, 22, 23]. These compilers assist the programmer in generating an efficient parallel implementation. To make suggestions for the design of a parallel implementation, e.g. a selection of a data distribution, the compiler has to have access to a powerful performance prediction tool to estimate the effects ....

I. Foster and K.M. Chandy. Fortran M: A Language for Modular Parallel Programming. Journal of Parallel and Distributed Computing, 1995.

....are multidisciplinary applications, in which di erent parts are developed independently as data parallel programs and interact at a coarser task level. This has caused a growing interest in models and systems which allow the expression (and the exploitation) of both task and data parallelism [2, 9, 16, 21 23]. In the general case, integration of task and data parallel models is rather dicult as these models tend to enforce di erent programming and implementation styles [2] However, most of the applications bene ting from a mixed style shows a fairly static and predictable structure [7, 10, 18] and ....

I. T. Foster and K. M. Chandi. Fortran M: a language for modular parallel programming. Journal of Parallel and Distributed Computing, 26(1):24-35, Apr. 1995.

....message passing represents a very low level for expressing interprocess communication. Message plumbing is hardwired into codes, requiring a significant recoding effort if communication patterns change. Furthermore, there is no explicit support for data parallelism. 7 3.2. 2 Fortran M Fortran M [9] extends Fortran 77 with a set of features that support message passing, according to a strictly enforced discipline. Fortran M processes program modules encapsulating data and code that are executed concurrently can be combined via channels; each channel establishes a one to one connection ....

I. T. Foster and K. M. Chandy. Fortran M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 25(1), 1995. 16

....endpoint and the methods used to communicate with that endpoint. These multimethod communication techniques have been implemented in the context of the Nexus multithreaded runtime system [15, 16] Nexus has been used to implement a variety of parallel languages and communication libraries [7, 13, 11], including the MPI implementation used extensively in the I WAY wide area computing experiment [10] We use Nexus to study the performance of alternative approaches to the implementation of various multimethod communication structures. We conclude with a case study in which our multimethod ....

I. Foster and K. M. Chandy. Fortran M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 25(1), 1994. 15

....vice versa. Implementation. We have developed for this archetype an implementation consisting of program transformation guidelines, together with a code skeleton and an archetype specific library of communication routines. The code skeleton and library are Fortran based, with versions in Fortran M [10], Fortran with p4, and Fortran with NX. The implementation is described in detail in [19] 4.3 Parallelization strategy As noted, in most respects our target application fits the pattern of the mesh archetype. The near field calculations are a perfect example of this archetype and thus can be ....

I. T. Foster and K. M. Chandy. FORTRAN M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 26(1):24--35, 1995.

....error min dgen diff diff diff error error error min dgen diff error Figure 12: Stereo task graph partitioned into 4 modules number of LATENCY THROUGHPUT modules msec frame frames sec 1 112 9.0 4 84 12.0 6 120 8. 3 Figure 13: Performance of 256 Theta 100 stereo partitioned different ways framework [5, 6, 14]. There is also a large body of literature on, partitioning, load balancing and scheduling of parallel programs [2, 11] We have addressed the specific partitioning and mapping issues that arise when task and data parallelism are combined in a parallelizing compiler, based on the characteristics ....

....partitioning and mapping issues that arise when task and data parallelism are combined in a parallelizing compiler, based on the characteristics of computations in the program. Our system is part of the Fx compiler that exploits task and data parallelism. Our approach is different from Fortran M [6] in the sense that data and task parallelism are strongly integrated; both are expressed and compiled uniformly using a sequential Fortran program and additional directives. An alternate approach, taken in Jade [9] is to express all parallelism as coarse grain tasks, and make scheduling decisions ....

FOSTER, I., AND CHANDY, K. Fortran M: A language for modular parallel programming. Tech. Rep. MCS-P3270992, Argonne National Laboratory, June 1992.

....We believe that our work is complementary to that of Anderson et al. Flick can leverage the benefits of an efficient transport system to produce optimized communication stubs. For parallel applications, there are a large number of specialized programming languages such as CC [7] Fortran M [16], and Split C [9] In most of these cases the language handles marshaling and unmarshaling of parameters. However, it is our belief that the techniques used by Flick, and possibly even its code, could be incorporated into the compilers for these languages to substantially reduce presentation layer ....

I. T. Foster and K. M. Chandy. Fortran M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 25(1), Feb. 1995.

....on the necessary communication and on the resulting performance. For many applications the exploitation of both task and data parallelism leads to more efficient parallel programs. Therefore, there is an increasing interest in mixed programming models, including the research on Fx [9] Fortran M [7], and Paradigm [3] We investigate a special class of applications exhibiting two levels of potential parallelism: method parallelism and system parallelism. The corresponding programs can be conceived as collections of modules which are connected by input output relations of data. A module ....

I. Foster and K.M. Chandy. Fortran M: A Language for Modular Parallel Programming. Journal of Parallel and Distributed Computing, 1995.

....to map distributed data structures onto this virtual machine, and the ability to define procedures that will run on each processor of the virtual machine concurrently. Process management and interprocessor communications are handled automatically by the compiler. 4.7. 2 Fortran M Fortran M [42] was developed at Argonne National Laboratory and is a language for modular parallel programming. It provides extensions to Fortran 77 to support message passing and includes the following features for providing modularity (programs are constructed using explicitly declared communication channels ....

I. Foster and K. M. Chandy. Fortran M: A language for modular parallel programming. J. Parallel and Distributed Computing, 25(1), 1995.

....global data that has to be shared between the processes and is realized by means of virtual memory. To regulate the data access the programmer has to make use of standard low level synchronization methods like semaphores and atomic updates making deadlocks possible. The approach taken by Fortran M [9] is based on explicit message passing. The programmer can declare processes which can exchange data through typed channels. These channels connect two arbitrary processes, again enabling deadlocks. A more static approach for the communication has been chosen within FX [13] where tasks can ....

I.T.Foster and K.M. Chandy. Fortran M: A Language for Modular Parallel Programming, June 1992.

.... ranging from signal processing to multidisciplinary problems for which an adequate load balance cannot be achieved on the basis of the SPMD model alone. Language features are needed for the explicit creation and manipulation of asynchronous tasks. Current proposals in that direction include [45, 17]. Another weakness of current systems is the lack of explicit, structured knowledge. While there is a significant body of knowledge in these systems, it is mostly hardwired and unstructured, scattered across many thousands lines of code, and thus not explicitly accessible to user or system ....

I.T.Foster and M.Chandy. Fortran M: A Language for Modular Parallel Programming. Technical Report, Department of Computer Science, California Institute of Technology (June 1992).

....targets runtime systems that provide distributed shared data or distributed shared objects. Nexus [14] is a distributed thread library that provides a virtual machine layer similar to Panda. Nexus was designed to be used as a back end for concurrent programming languages, e.g. CC and FortranM [13]. It provides support for creating threads on a processor, for specifying synchronization among the threads of a node, and for communication among threads through the active message paradigm. 2.2 The Component Approach The component approach, illustrated in Figure 2, attempts to separate ....

I. Foster and K. M. Chandy. Fotran M: A Language for Modular Parallel Programming. Journal of Parallel and Distributed Computing, 25(1), 1995.

....archetype, including: ffl Fortran M (1D, 2D, 3D, and 3D without a host process) ffl Fortran plus NX (2D and 2D without a host process) In general, the user level specifications of all implementations are the same. Where they differ, we describe here the Fortran M 3D implementation. Fortran M [2, 1] is a Fortran variant available via anonymous FTP from Argonne National Labs. 2.1 Parameter definitions The application programmer supplies PARAMETER definitions for the following: ffl The dimensions of the (N dimensional) grid. ffl The width of the ghost boundary. This width cannot be ....

....ipx, ipy, ipz, inname, outname) In no host process implementations, the argument iptype is omitted. 3 Compiling and linking Fortran M version This section describes how to compile and link a program based on the Fortran M implementation. It assumes that the Fortran M compiler (fm) [2, 1] is installed. To compile and link a program based on the Fortran M implementation: ffl Compile the main program file the file containing the archetypesupplied program skeleton plus the user supplied code for the programs that execute in the host and grid processes. ffl Compile any ....

I. T. Foster and K. M. Chandy. FORTRAN M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 26(1):24--35, 1995.

....defined and implemented using this notation and preserved as reusable templates. In PCN, logical variables are used as channel variables for communication synchronisation based on their single assignment feature. Based on the PCN, a class of parallel languages has been proposed such as Fortran M [7], where Fortran is taken as the base language for sequential programming. Unfortunately, this clean methodology is still quite low level. Thus, high level compositions have to be programed with lower level primitives concerned with process activations, interprocess communications and process ....

I. Foster and K. M. Chandy. Fortran M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 25(1), 1995.

....be skipped by a generalized runtime system. In addition, the compiler could recognize and exploit further potential parallelism like intra SDA parallelism which is a major topic for future research. Related coordination approaches that make use of a compiler are, for example, Fx [17] FortranM [10], or Orca [2] Fx does not need any runtime support since it has a quite simple tasking model similar to the tasking facilities in HPF 2 (in fact, HPF 2 took over many ideas introduced in Fx) Fortran M is based upon Nexus which fits perfectly into the Fortran M model. However, Fortran M is a ....

I. Foster and K.M. Chandy. Fortran M: A Language for Modular Parallel Programming. Journal of Parallel and Distributed Computing, Vol. 26, 1995.

....prediction techniques [9, 17, 19] and parallelizing compilers. In the following, we compare our approach to similar approaches in the area of parallelizing compilers. Several research groups working on parallelizing compilers have included support to combine task and data parallelism. Fortran M [11, 12] allows the creation of processes which can communicate with each other by predefined channels and which can be combined with HPF for a mixed task and data parallel execution. In contrast to the Fortran M approach, our approach is more directed towards the combination of fine and medium grain ....

I. Foster and K.M. Chandy. Fortran M: A Language for Modular Parallel Programming. Journal of Parallel and Distributed Computing, 1995.

....The body of this encapsulation can contain only loops and subroutine calls. Assigning processors to the task is done using processor and origin directives. Since the model of parallelism is processor based, different co executing tasks cannot be scheduled on the same set of processors. Fortran M[8] adds task parallel extensions to Fortran 77. Here, a process models a task and concurrent processes communicate through message passing. The runtime system of Fortran M is based on a thread system called Nexus[9] Message passing is done over channels rather than between threads. A process ....

Ian Foster and Mani Chandy. Fortran M: A language for modular parallel programming. Technical Report MCS-P327-0992, Argonne National Laboratory, June 1992.

....module is implemented on a number of nodes. Parallel programs can be composed from independent modules, which are themselves parallel programs. Such parallel programs may arise when a parallelizing compiler exploits task parallelism[6] or such programs may be written directly by a human programmer[5, 4, 1]. At runtime, these modules must interface with each other by exchanging messages. Since the modules have been created separately, and we want to compose modules in a modular fashion without undue constraints, there are additional requirements on creating and consuming these interface messages. ....

I. T. Foster and K. M. Chandy. Fortran M: A Language for Modular Parallel Programming. Technical Report ANL, Argonne National Laboratory, 1993.

....languages libraries We can use the transformation of the preceding section to transform programs in the subset par model into programs in any language that supports (1) multiple address space parallel composition with (2) single sender single receiver message passing. Examples include Fortran M [40] (which supports multiple address space parallel composition via process blocks and single sender singlereceiver message passing via channels) and MPI [58] which assumes execution in an environment of multiple address space parallel composition and supports single sender single receiver ....

....of an archetype implementation. 7.2.1.4 Implementation We have developed for the mesh spectral archetype an implementation consisting of a code skeleton and an archetype specific library of communication routines for the operations described in Section 7.2.1. 3, with versions based on Fortran M [40] and Fortran with MPI [58] The implementation is described in detail in [34] The Fortran M version has been used to run applications on the IBM SP and on networks of Sun workstations; the MPI version has been used to run applications on the IBM SP and on networks of Sun and Pentium based ....

[Article contains additional citation context not shown here]

I. T. Foster and K. M. Chandy. FORTRAN M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 26(1):24--35, 1995.

....encapsulated in the redistribution routine. This algorithm has been implemented on top of a general mesh spectral archetype implementation (consisting of a code skeleton and an archetype specific library of communication routines) The archetype in turn has been implemented in both Fortran M [23] and Fortran with MPI [29] The Fortran M version has been used to run applications on the IBM SP and on networks of Sun workstations; the MPI version has been used to run applications on the IBM SP and on networks of Sun and Pentium based workstations. Figure 12 shows speedups of the MPI version ....

I. T. Foster and K. M. Chandy. FORTRAN M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 26(1):24--35, 1995.

....on the condition variable, notifying the initial thread that the remote operation has completed. Because remote service requests are unidirectional, the CC compiler can detect when a return value is not required and optimize out the return remote service request. 4. 2 Compiling FM Using Nexus FM [13, 15] is a small set of extensions to Fortran 77 for task parallel programming. FM is designed to support both the modular construction of large parallel programs and the development of libraries implementing other programming paradigms. For example, in a joint project with Syracuse, such a library has ....

Ian Foster and K. Mani Chandy. Fortran M: A language for modular parallel programming. J. Parallel and Distributed Computing, 1994. to appear.

....structure, improving productivity. Use of familiar tools. The methods and techniques provided by an archetype are also languageindependent; the code libraries associated with an archetype can be implemented using different parallel languages and libraries (for example, Fortran M [12] and PVM [3] Stepwise refinement allows much of the work of developing a parallel program to be done in a sequential environment, using familiar tools and techniques; the choice of a particular parallel language or library can be deferred until the last step of the refinement process. This ....

....attention to a few common types that suffice to cover many applications. As an additional simplification, we require that all group to group communication take place over explicitly defined typed communication paths or channels, analogous to the typed process to process channels of Fortran M [12]. A group to group channel is typed with the distribution and restriction functions of both the sender and the receiver as well as the interface mapping. This information makes it possible to translate each group to group channel into a set of process to process communication paths, as illustrated ....

[Article contains additional citation context not shown here]

I. T. Foster and K. M. Chandy. FORTRAN M: A language for modular parallel programming. Journal of Parallel and Distributed Computing, 1994. To appear.

No context found.

I. Foster and K.M. Chandy. Fortran M: A Language for Modular Parallel Programming. Journal of Parallel and Distributed Computing, 25(1):24--35, April 1995.

No context found.

I. Foster and K.M. Chandy. Fortran M: A Language for Modular Parallel Programming. Journal of Parallel and Distributed Computing, 1995.

No context found.

I. Foster and K. M. Chandy. Fortran M : a language for modular parallel programming. In Journal of Parallel and Distributed Programming.

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