B. Kruatrachue and T. Lewis. Grain Size Determination for Parallel Processing. IEEE Software, January 1988.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....is required between tasks. In an environment where the hardware timing parameters are known before the simulation is executed, one approach is to start with the finest grain representation possible and combine fine grain tasks together to create coarser grained tasks by employing grain packing[8] techniques. Such techniques make use of the targeted topology and selected task allocation methodology to determine the granularity of each task. This allows much of the work to be automated, but the resulting representation loses much of the recognizable structure of the system. Another approach ....

....resent the system model. It can be executed in a manner that reserves the re resentation of tasks s ecified by the user in the Model Descri tion File or it can be instructed to automatically construct tasks via a number of automatic methodologies such as the grain acking methods discussed earlier [8]. In some cases, the task gra h creater may actually execute sequential simulations in order to obtain task weighting data, or it may be instructed to use rede 15 Simulation, Vol. 65, N, 3, pp. 191 205, September 1995. termined weightings on computational constructs to compute the aggregate ....

B. Kruatrachue and T. Lewis, "Grain Size Determination for Parallel Processing," IEEE Software, Vol. 5, N[AM[wN[wNRE[OwTOO . 23-31. Simulation, Vol. 65, N, 3, pp. 191-205, September 1995.

....the earliest start time is selected to accommodate this node. Most of the reported scheduling algorithms are based on this concept of employing variations in the priority assignment methods, such as HLF (Highest level First) LP (Longest Path) LPT (Longest Processing Time) and CP (Critical Path) [1, 24, 15]. In the following we review some of contemporary static scheduling algorithms, including MCP, DSC, DLS, and CPN methods. The Modi ed Critical Path (MCP) algorithm is based on the as late as possible (ALAP) time of a node [24] The ALAP time is de ned as TL (n i ) T critical level(n i ) ....

B. Kruatrachue and T. Lewis. Grain size determination for parallel processing. IEEE Software, pages 23-32, January 1988.

....the code based on graph manipulation facilities. HTGviz also guides the user through the process of generating valid and efficient parallel code for OpenMP applications. 1 Introduction Parallel programs are collections of tasks that could run serially or in parallel. The grain size problem [1, 9, 11] is how to determine the best trade off between task grain (amount of parallelism) and overhead. Sources of overhead in a multiprocessor setting include scheduling, synchronization and communication overhead [7, 2] Partitioning techniques are necessary to execute parallel programs at appropriate ....

Kruatrachue, D., Lewis, T.: Grain size determination for parallel processing. IEEE Software (1988) 23--32

....into tasks, we need to determine the best tradeoff between parallelism and overhead. The grain size problem is related to the max min problem, because there is a trade off between parallelism (fine grain) and communication (large grain) This problem has been extensively studied, see for example [3, 13, 16]. Therefore, to grasp the optimal structure for a given program task graph, we must relies on the user knowledge. During compilation inside a GPE session, the user may interact with different views of the program task graph and textual representations: he (she) gets information like concurrency ....

Kruatrachue D. and Lewis T. Grain size determination for parallel processing. IEEE Software (1988), 23--32.

....if the two tasks are assigned to different processors. Given a PDG, the graph is partitioned into appropriately sized grains which are assigned to processors of a parallel machine. The partitioning and assignment known as the scheduling problem. The problem is also called grain size determination [2], the clustering problem [3, 4] and internalization pre pass [1] The partitioning scheduling problem is intractable, and heuristics are required to find sub optimal solutions. As a result, there are no performance guarantees for scheduling heuristics for general graphs. Many researchers have ....

....has received the most attention. A taxonomy of these techniques as well as a comparison of four specific heuristics can be found in the work of Gerasoulis and Yang [8] In this paper we include experimental results from two critical path algorithms, DSC and MCP. ffl List scheduling heuristics [2, 7, 12, 13, 14, 15, 16]: These algorithms assign priorities to the tasks and schedule them according to a list priority scheme. For example, a high priority might be given to a task with many heavily weighted incident edges or to a task whose neighbors have already been scheduled. Extending the list scheduling heuristic ....

[Article contains additional citation context not shown here]

Kruatrachue B. and Lewis T. Grain Size Determination for Parallel Processing. IEEE Software, pages 23--32, Jan 1988.

....the graph is partitioned into appropriately sized grains which are assigned to processors of a parallel machine. A good assignment will shorten the execution time of the program. The partitioning and assignment is called the scheduling problem. The problem is also known as grain size determination [12], the clustering problem [11, 24] and internalization pre pass [21] The problem is important because solution methods can be used to generate efficient parallel programs. The partitioning scheduling problem is intractable, and heuristics are required to find sub optimal solutions. In addition, ....

....the most attention. A taxonomy of these techniques as well as a comparison of four specific heuristics can be found in the work of Gerasoulis and Yang [4] In this paper we include experimental results from three critical path algorithms, DSC, Linear Clustering and MCP. List scheduling heuristics [1, 10, 12, 13, 14, 18, 20]: These algorithms assign priorities to the processes and schedule them according to a list priority scheme. For example, a high priority might be given to a task with many heavily weighted incident edges or to a task whose neighbors have already been scheduled. Extending the list scheduling ....

[Article contains additional citation context not shown here]

Kruatrachue, B. and Lewis, T. Grain Size Determination for Parallel Processing. IEEE Softw., Jan. 1988, pp. 23-32.

....assigned to different processors. Given a PDG, the graph is partitioned into appropriately sized groups of nodes (grain) which are assigned to processors of a parallel machine. The partitioning and assignment are known as the scheduling problem. The problem is also called grain size determination [2], the clustering problem [3, 4] and internalization pre pass [1] The partitioning scheduling problem is intractable, and heuristics are required to find sub optimal solutions. As the result, there are no performance guarantees for scheduling heuristics for general graphs. Many researchers have ....

....the most attention. A taxonomy of these techniques as well as a comparison of four specific heuristics can be found in the work of Gerasoulis and Yang [8] In this paper we include experimental results from four critical path algorithms, DSC, DCP, MCP and HU. 1 ffl List scheduling heuristics [2, 7, 14, 15, 16, 17, 18]: These algorithms assign priorities to the tasks and schedule them according to a list priority scheme. For example, a high priority might be given to a task with many heavily weighted incident edges or to a task whose neighbors have already been scheduled. Extending the list scheduling heuristic ....

[Article contains additional citation context not shown here]

B. Kruatrachue and T. Lewis, "Grain Size Determination for Parallel Processing," IEEE Software, pages 23--32, Jan 1988.

....if the two tasks are assigned to different processors. Given a PDG, the graph is partitioned into appropriately sized grains which are assigned to processors of a parallel machine. The partitioning and assignment known as the scheduling problem. The problem is also called grain size determination [2], the clustering problem [3, 4] and internalization pre pass [1] The partitioning scheduling problem is intractable, and heuristics are required to find sub optimal solutions. As the result, there are no performance guarantees for scheduling heuristics for general graphs. Many researchers have ....

....attention. A taxonomy of these techniques as well as a comparison of four specific heuristics can be found in the work of Gerasoulis and Yang [8] In this paper we include experimental results from three critical path algorithms, DSC, Linear Clustering, and MCP. ffl List scheduling heuristics [2, 7, 12, 13, 14, 15, 16]: These algorithms assign priorities to the tasks and schedule them according to a list priority scheme. For example, a high priority might be given to a task with many heavily weighted incident edges or to a task whose neighbors have already been scheduled. 1 Extending the list scheduling ....

[Article contains additional citation context not shown here]

Kruatrachue B. and Lewis T. Grain Size Determination for Parallel Processing. IEEE Software, pages 23--32, Jan 1988.

....Many existing data parallel implementations are severely limited in their scope of partitioning strategies (e.g. limiting all aggregates to simple block partitioning) and decide upon strategy according to simplistic schemes with little consideration of aggregate use. Considerable research (e.g. [10, 18, 13]) has shown that optimally partitioning data across a distributed memory machine should be determined through complex analysis of the entire program. Such analysis may generate quite unusual partitioning schemes for data items used in irregular fashions. We endeavor to place few restrictions upon ....

B. Kruatrachue and T. Lewis. Grain size determination for parallel processing. IEEE Software, pages 23--32, January 1988.

....too small, it may happen that the costs are larger than the benefits in their parallel execution. This makes it desirable to devise a method whereby the granularity of parallel goals and their number can be controlled. Granularity control has been studied in the context of traditional programming (Kruatrachue and Lewis 1988, McGreary and Gill 1989) functional programming (Huelsbergen 1993, Huelsbergen et al. 1994) and also logic programming (Kaplan 1988, Debray et al. 1990, Zhong et al. 1992, Debray and Lin 1993) The benefits from controlling parallel task size will obviously be greater for systems with greater ....

Kruatrachue, B., Lewis, T. (1988). Grain Size Determination for Parallel Processing. IEEE Software, January.

.... in order to minimize the completion time on the parallel computer system [10] 15] 17] 45] 55] 65] While job scheduling requires dynamic run time scheduling that is not a priori decidable, the scheduling and mapping problem can be addressed in both static [4] 7] 9] 18] 67] 39] [42], 43] 44] 50] 51] 52] 56] 58] 61] 62] 70] as well as dynamic contexts [2] 3] 37] When the structure of the parallel program in terms of its task execution times, task dependencies, task communications and synchronization, is known a priori, scheduling can be accomplished ....

....depend. In this section, we discuss two previously proposed algorithms and discuss their merits and weaknesses. These algorithms assume the availability of unlimited number of processors. Using duplication in static scheduling is a relatively unexplored research topic. Kruatrachue and Lewis [42] have proposed one such scheduling algorithm, called Duplication Scheduling Heuristic (DSH) Another algorithm, called Bottom up Top down Duplication Heuristic (BTDH) has been recently proposed by Chung and Ranka [18] These two algorithms are briefly described below. 2.6.1 The DSH Algorithm ....

B. Kruatrachue and T.G. Lewis, "Grain Size Determination for Parallel Processing," IEEE Software, pp. 23-32, Jan. 1988.

....overhead for parallelism over the actual execution time. To avoid this, the grain size of each parallel task has to be taken into account. There has been a lot of work on estimating grain size of parallel tasks in various contexts (IAP [31, 9] functional programming [17] and imperative languages [20]) While granularity control appears to be sufficient to tackle the efficiency issues in those forms of parallelism where parallel computations are independent (i.e. IAP and ORP) it is not sufficient to guarantee good execution efficiency of parallelism in the case of DAP. Consider two subgoals ....

B. Kruatrachue and T. Lewis. Grain Size Determination for Parallel Processing. IEEE Software, January 1988.

....derivation of task periods [18, 26, 46] By means of task replication, the performance of the application can be greatly improved. Much work in recent years has shown that replication may be needed for fault tolerance [21, 43] or schedulability [9, 18] reasons, or for minimizing schedule length [36, 11, 3, 37]. The purpose of task clustering is to provide directions to the subsequent task allocation stage as regards the task to processor assignments. Task clustering has, primarily, been used for guaranteeing that certain tasks that exchange large amounts of data are assigned to the same processor in ....

B. Kruatrachue and T. Lewis. Grain size determination for parallel processing. IEEE Software, 5(1):23--32, January 1988.

....has been shown to be NP complete [3] The general problem of task scheduling has been extensively studied, mainly for homogeneous systems. Various heuristics have been proposed, including list algorithms [4, 11, 12, 13, 20] multi step algorithms [14, 15, 22] duplication based algorithms [7, 2, 1], genetic algorithms [18] algorithms using local search [21] bin packing [19] or graph decomposition [6] Within all these approaches, list scheduling has been shown to have a good cost performance trade off, as considering its low cost, the performance is still very good [8, 13, 12] The ....

B. Kruatrachue and T. G. Lewis. Grain size determination for parallel processing. IEEE Software, pages 23--32, Jan. 1988.

....use, because the required number of processors is not usually available. Hence, their application is typically found within the multi step scheduling method for a bounded number of processors [7, 8, 11] Scheduling for a bounded number of processors can be done either using duplication (e.g. DSH [5] or CPFD [1] or without duplication (e.g. MCP [10] ETF [4] or DSC LLB [7, 11] Duplicating tasks results in better performance but significantly increases cost compared to non duplicating heuristics. Within non duplicating heuristics, list scheduling algorithms obtain good performance at a low ....

B. Kruatrachue and T. G. Lewis. Grain size determination for parallel processing. IEEE Software, pages 23--32, Jan. 1988.

....based scheduling, critical tasks are redundantly scheduled to more than one machines in order to reduce the number of inter task communication operations. The start times of the succeeding tasks are also reduced. There have been many duplication approaches suggested in the literature [1] [7], 13] 14] 15] 17] However, all these methods are designed for homogeneous parallel architectures. Furthermore, the previous approaches are all evaluated based on simulations rather than using real applications with a parallelizing compiler. In our proposed approach, the task duplication ....

B. Kruatrachue and T.G. Lewis, "Grain Size Determination for Parallel Processing," IEEE Software, vol. 5, no. 1, pp. 23-32, Jan. 1988.

....at the level of the scheduling algorithms for unbounded number of processors. However, the schedule lengths increase up to 36 compared to a list scheduling algorithm like MCP [9, 10] Scheduling for a bounded number of processors can be also done in one step, either using duplication (e.g. DSH [8], BTDH [3] or CPFD [2] or without duplication (e.g. MCP [14] ETF [7] DLS [12] or HLFET [1] Duplicating tasks results in better scheduling performance but signi cantly increases scheduling cost. Non duplicating list scheduling heuristics have a lower complexity and still obtain good ....

B. Kruatrachue and T. G. Lewis. Grain size determination for parallel processing. IEEE Software, pages 23-32, Jan. 1988.

....of processors [8, 9] Alternatively, scheduling for a bounded number of processors can be performed in a single step. Using this one step approach, the results are usually better, but at a higher cost. Scheduling for a bounded number of processors can be done either using duplication (e.g. DSH [4], BTDH [2] or CPFD [1] or without duplication (e.g. MCP [11] ETF [3] DLS [10] or FCP [7] Duplicating tasks results in better scheduling performance but significantly increases scheduling cost. Nonduplicating task heuristics have a lower complexity and still obtain good schedules. However, ....

B. Kruatrachue and T. G. Lewis. Grain size determination for parallel processing. IEEE Software, pages 23--32, Jan. 1988.

No context found.

B. Kruatrachue and T. Lewis. Grain Size Determination for Parallel Processing. IEEE Software, January 1988.

No context found.

B. KRUATRACHUE and T. LEWIS. Grain size determination for parallel processing. IEEE Software, 5(1):23--32, January 1988.

No context found.

B. Kruatrachue and T. Lewis. Grain Size Determination for Parallel Processing. IEEE Software, January 1988.

No context found.

B. Kruatrachue and T. Lewis. Grain Size Determination for Parallel Processing. IEEE Software, 5(1):23--32, January 1988.

No context found.

B. Kruatrachue and T. G. Lewis. Grain size determination for parallel processing. IEEE Software, pages 23--32, Jan. 1988.

No context found.

B. Kruatrachue and T. G. Lewis. Grain size determination for parallel processing. IEEE Software, pages 23--32, Jan. 1988.

No context found.

B. Kruatrachue and T. Lewis. Grain Size Determination for Parallel Processing. IEEE Software, vol. 5(no. 1):23--32, 1988.

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