Clemens H. Cap and Volker Strumpen. Efficient Parallel Computing in Distributed Workstation Environments. Parallel Computing, 19:1221--1234, 1993.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....computation among the executive processes and collects results. The application independent code of the executive process sets up the communication links and supports message passing and dynamic load balancing. The application dependent part of the code performs the arbitrary size subtasks. In [15], Cap and Strumpen address the issue of utilizing heterogeneous net 22 works of workstations with constantly changing load situation for parallel computa tions. In frequently synchronizing computations, the throughput of a heterogeneous network is reduced to that of the slowest workstation. They ....

....adapt to the computing environment and analyze their performance. 4.1 Nondedicated Workstations as Transient Processors Processors in a network of workstations (NOW) are often underutilized. Sev eral studies indicate that a large number of workstations in a NOW are idle at any given time [15, 38, 52]. Arpaci et al. [4] report that, although the set of idle machines changes over time, the total number of idle machines stays relatively con stant. Our objective is to use the idle workstations to run additional jobs. There have been several systems that attempt to make use of idle workstations to ....

Clemens H. Cap and Volker Strumpen. Efficient Parallel Computing in Distributed Workstation Environments. Parallel Computing, pages 1221-1234, 1993.

....for a cluster of 25 workstations, mayhaveapplications to future parallel computers as well. 1.3. 1 Comparison with other work in parallel computing The suitabilityoflocal interaction algorithms for parallel computing on a cluster of workstations has been demonstrated in previous works, suchas[7], 9] and elsewhere. Cap Strumpen [7] presentthePARFORM system and simulate the unsteady heat equation using explicit finite differences. Chase et al. 9] presenttheAMBER system, and solve Laplace s equation using SuccessiveOver Relaxation. The present work emphasizes, and clarifies further the ....

....mayhaveapplications to future parallel computers as well. 1.3.1 Comparison with other work in parallel computing The suitabilityoflocal interaction algorithms for parallel computing on a cluster of workstations has been demonstrated in previous works, suchas[7] 9] and elsewhere. Cap Strumpen [7] presentthePARFORM system and simulate the unsteady heat equation using explicit finite differences. Chase et al. 9] presenttheAMBER system, and solve Laplace s equation using SuccessiveOver Relaxation. The present work emphasizes, and clarifies further the importance of local interaction methods ....

[Article contains additional citation context not shown here]

C. H. Cap and V. Strumpen. Efficientparallel computing in distributed workstation environments. Parallel Computing,19(11):" 34, 1993. 251

....employ static job scheduling strategies. Massively parallel supercomputers such as the Cray Research T3D or the Thinking Machines CM5, for example, either dedicate themselves to a single user at a time or gang timeshare within fixed size partitions [72] Systems such as Charm [91] the Parform [21], PVM Hence [96] and others [29, 42, 44, 97] support parallel computing on a network of workstations. In these systems, the set of machines on which the program runs is chosen statically by the user. Distributed operating systems [30, 82, 98, 99] and other systems [32, 40, 68, 74, 79, 110] ....

Clemens H. Cap and Volker Strumpen. Efficient parallel computing in distributed workstation environments. Parallel Computing, 19:1221--1234, 1993.

....chances to use the resources of the slower nodes. 1.1 Applications There are two types of parallelism to be exploited on multicomputers: task parallelism and data parallelism. While many load balancing algorithms focus on task parallelism to minimize the average response time of tasks[7] 9] 10][11], our goal is to exploit data parallelism within a task, that is, to minimize the total execution time of a program. Data parallel applications can be classified according to their characteristics. When it comes to data dependency, they can be categorized as independent, adjacently dependent and ....

Clemens H. Cap and Volker Strumpen. "Efficient Parallel Computing in Distributed Workstation Environemnts". Parallel Computing,Vol.19, pp.1221-1234, 1993.

....time. Workstations are only intermittently idle as a rule. A system which can utilize intermittently idle workstations can make computing on networked workstations very cost effective. In fact, for sequential or semi parallel (i.e. multiple tasks with no inter dependency) jobs, there are systems[2, 14, 18, 11, 25, 28, 44, 48, 52, 53] to utilize idle or under utilize workstations effectively. For parallel computation, there are also systems[2, 11, 25, 44] In the last several years, there has been a proliferation of commercial and research prototype parallel software systems on networks of workstations. Popular systems ....

....workstations very cost effective. In fact, for sequential or semi parallel (i.e. multiple tasks with no inter dependency) jobs, there are systems[2, 14, 18, 11, 25, 28, 44, 48, 52, 53] to utilize idle or under utilize workstations effectively. For parallel computation, there are also systems[2, 11, 25, 44]. In the last several years, there has been a proliferation of commercial and research prototype parallel software systems on networks of workstations. Popular systems include Linda[13] PVM[60, 24] MPI[29] and Express. Unfortunately, few support fault tolerance or utilization of idle ....

[Article contains additional citation context not shown here]

Clemens H. Cap and Volker Strumpen. Efficient parallel computing in distributed workstation environments. Parallel Computing, 19:1221--1234, 1993.

....: pattern = f string g : interface = null : g 3.2.3 Solution 3: Dynamic load balancing When load balance cannot be reached through the above load distribution methods, tasks should migrate. Many dynamic load balancing algorithms have been devised for such an efficient migration [6, 10, 17, 18, 23]; they are characterized by the following parameters which distinguish them. Load balancing algorithms can be fine tuned when programmers can change those factors conveniently. ffl Topology: Topology determines the shape of task migration paths. A fully connected topology provides a way to gain ....

Clemens H. Cap and Volker Strumpen. Efficient parallel computing in distributed workstation environments. Parallel Computing, Vol. 19:1221--1234, 1993.

....local work as possible before a connection to the LSS is made. ffl P4. External processing capabilities: networks of workstations (NOWs) are an attractive alternative to make available the idle cycles, unused main memory, and disk space of a collection of workstations, to parallel programs [1] [2]. An LSS should be capable of registering idle CPU cycles at user facilities and schedule computations using these cycles. There should also be a mechanism by which users of an LSS could pay for part of their usage through cycles. ffl P5. Separation of Functions: the functions provided by an LSS ....

Cap, C. H. and V. Strumpen, Efficient Parallel Computing in Distributed Workstation Environments , Parallel Computing, 19:1221-1234, 1993.

....: pattern = f string g : interface = null : g 3.8 Solution 3: Dynamic load balancing When load balance cannot be reached through the above load distribution methods, tasks should migrate. Many dynamic load balancing algorithms have been devised for such an efficient migration [6, 10, 17, 18, 23]; they are characterized by the following parameters which distinguish them. Load balancing algorithms can be fine tuned when programmers can change those factors conveniently. ffl Topology: Topology determines the shape of task migration paths. A fully connected topol ogy provides a way to gain ....

Clemens H. Cap and Volker Strumpen. Efficient parallel computing in distributed workstation environments. Parallel Computing, Vol. 19:1221-- 1234, 1993.

.... processor will have T= n 1) Therefore, if new n n 1 min , the execution time of (n 1) processors cluster is T (n 1)new , which is longer than that of n processors One may want to circumvent this problem by allocating tasks according to the known computing power of each processor [9, 3]. However, their methods were static, thus of limited usefulness. Dynamic loop scheduling methods can deal with more general cases, but the centralized nature of the methods the central processor that generates sub tasks has to manage all other processors may cause a bottleneck in a ....

Clemens H. Cap and Volker Strumpen. Efficient parallel computing in distributed workstation environments. Parallel Computing, Vol. 19:1221-- 1234, 1993.

....(NOW) has become the rational choice to perform heavy parallel and distributed computations. In order to find the distributed resources that match at best the distributed processes requirements, many research efforts have been done (and are still done) on dynamic load balancing (DLB) algorithms [6, 16, 8, 12, 7, 10, 4]. Out of the many problems one is faced with the DLB algorithm design, we are especially interested in the definitions and meanings of the load indicators : clearly, the best our knowledge of the global system state and behavior is, the wiser our workload sharing or balancing policies should be. ....

....or balancing policies should be. The main difficulties to consider with NOWs are firstly, their multi user and multi programming purpose and secondly, their heterogeneity. A quite usual solution used to face the first difficulty, consist in building DLB algorithms that look for idle workstations [10, 4] : such algorithms are intended not to disturb the users owners of the workstations and are practically quite easy to implement, given the boolean nature of the load indicator (idle or busy) But this kind of algorithms also have noticeable drawbacks. In particular, it does not allow to fully ....

C. H. CAP and V. STRUMPEN. Efficient parallel computing in distributed workstation environnements. Parallel Computing, 19:1221--1234, 1993.

....a cluster of 25 workstations, may have applications to future parallel computers as well. 1.3. 1 Comparison with other work in parallel computing The suitability of local interaction algorithms for parallel computing on a cluster of workstations has been demonstrated in previous works, such as [7], 9] and elsewhere. Cap Strumpen [7] present the PARFORM system and simulate the unsteady heat equation using explicit finite differences. Chase et al. 9] present the AMBER system, and solve Laplace s equation using Successive Over Relaxation. The present work emphasizes, and clarifies further ....

....applications to future parallel computers as well. 1.3.1 Comparison with other work in parallel computing The suitability of local interaction algorithms for parallel computing on a cluster of workstations has been demonstrated in previous works, such as [7] 9] and elsewhere. Cap Strumpen [7] present the PARFORM system and simulate the unsteady heat equation using explicit finite differences. Chase et al. 9] present the AMBER system, and solve Laplace s equation using Successive Over Relaxation. The present work emphasizes, and clarifies further the importance of local interaction ....

[Article contains additional citation context not shown here]

C. H. Cap and V. Strumpen. Efficient parallel computing in distributed workstation environments. Parallel Computing, 19(11):1221--1234, 1993. 252 BIBLIOGRAPHY 253

....applications with cooperative tasks. This includes the updating of routing tables to reflect a task s new location. Performance measurements [41] yielded almost full Ethernet communication bandwidth when migrating tasks on TCP. 4 Parform. The Parform system developed at the UniversityofZurich [8, 9], is a selfcontained parallel programming system. Applications are written in C with special parallelization extensions offered byParform. In a so called heterogeneous partitioning mode 1 ,theParform task mapper takes the individual load factors of the single processors into account. Tasks ....

....core images. Instead, the application code must be modified (e.g. by a preprocessor) to include calls to the runtime system at certain points where task migration is allowed. The user s object code is linked to the MARS runtime library which notifies 1 Note that Cap and Strumpen s [8, 9] heterogeneous workstation clusters are in fact homogeneous in the sense that they only include systems that are binary code compatible (e.g. Sparc1, Sparc2, SparcServer) Thus, core images can be migrated and restarted without further provision. 5 NW NW Checkpoint Send Information Send ....

C.H. Cap, V. Strumpen. Efficient parallel computing in distributed workstation clusters. Parallel Computing 19,11(1993), 1221--1234.

....9, 13, 17, 20, 22, 28, 29, 31] key components to achieving parallel speed up and high system throughput. More recently, performance studies of LAN and WAN based systems have shown the importance of controlling network communication for improving parallel or distributed application performance [5, 8, 18, 19, 25, 26]. The sensitivity of application performance to congestion effects is directly dependent upon the communication computation ratio and degree of synchrony. An application with a high communication computation ratio is prone to generate periods of concentrated congestion which can lead to ....

Clemens Cap and Volker Strumpen. Efficient parallel computing in distributed workstation environments. Parallel Computing, 19:1221--1234, 1993.

....machines. Over the years, a number of parallel programming languages and run time environments have been developed for distributed systems. Some of the earliest programming environments focus on coarse grain parallel applications; for example, p4 [108] Data parallel C [127] PVM [63] PARFORM [28], and MPI [161] Later systems use objects which can be migrated between idle nodes and invoked regardless of location; for example, Emerald [84] and CHARM [146] Finally, the most sophisticated systems, developed more recently, adjust the level of parallelism in the application to dynamically ....

Clemens H. Cap and Volker Strumpen. Efficient Parallel Computing in Distributed Workstation Environments. Parallel Computing, 19:1221--1234, 1993.

....migration [1, 9, 12, 18, 20] key components to achieving parallel speed up and high system throughput. More recently, performance studies of LAN and WAN based systems have shown the importance of controlling network communication for improving parallel and distributed application performance [2, 5, 13, 14]. The sensitivity of application performance to congestion effects is directly dependent upon the communication computation ratio and degree of synchrony. Distributed operating systems, for the most part, are written from scratch and are kernel based with a well defined interface to kernel ....

C. Cap and V. Strumpen. Efficient parallel computing in distributed workstation environments. Parallel Computing, 19:1221--1234, 1993.

....the processors among the parallel jobs. Time sharing disciplines use some form of round robin scheduling between the parallel jobs. A workstation cluster may also be considered as a parallel computer. A number of research activities have tried to exploit the computing power of such environments [1, 4, 5, 7]. A popular tool to implement parallel applications on workstation clusters is PVM (Parallel Virtual Machine) With PVM, each user maps her application statically onto nodes which are specified in a configuration file. But in most cases, workstation clusters are not used exclusively by one user ....

C.H. Cap and V. Strumpen. Efficient parallel computing in distributed workstation environments. Parallel Computing, 19(11):1221--1234, 1993.

.... [12] consider SIMD programs; ADM is not yet transparent and its portability is application based rather than platform based [8] UPVM and MPVM are oriented towards task parallelism [8] The first tool for irregular data distributions derived from networked computing platforms is The Parform [6] that efficiently supports heterogeneous data partitions. However, unlike ADD, The Parform requires 18 further investigations for generalising the dynamic load balancing strategies which are indispensable for computations on inherently variable platforms. 6. Experimental Results We evaluate the ....

C.H. Cap and V. Strumpen, Efficient parallel computing in distributed workstation environments, Parallel Computing 19 (1993) 1221-1234.

....steps in order to follow the fast moving acoustic waves. Thus, subsonic flow is a good problem for explicit methods. 1. 1 Comparison with other work The suitability of local interaction algorithms for parallel computing on a cluster of workstations has been demonstrated in previous works, such as [1], 2] and elsewhere. Cap Strumpen [1] present the PARFORM system and simulate the unsteady heat equation using explicit finite differences. Chase et al. 2] present the AMBER parallel system, and solve Laplace s equation using Successive Over Relaxation. The present work emphasizes and clarifies ....

....acoustic waves. Thus, subsonic flow is a good problem for explicit methods. 1.1 Comparison with other work The suitability of local interaction algorithms for parallel computing on a cluster of workstations has been demonstrated in previous works, such as [1] 2] and elsewhere. Cap Strumpen [1] present the PARFORM system and simulate the unsteady heat equation using explicit finite differences. Chase et al. 2] present the AMBER parallel system, and solve Laplace s equation using Successive Over Relaxation. The present work emphasizes and clarifies further the importance of local ....

[Article contains additional citation context not shown here]

C. H. Cap and V. Strumpen, "Efficient parallel computing in distributed workstation environments," Parallel Computing, vol. 19, no. 11, pp. 1221--1234, 1993.

....Today, many institutions have a LAN of workstations. Essentially being used in a interactive way, workstation stay idle for a long time. Some experiments have reported an average idle time rate of 90 [3] A network of workstations may be considered as a parallel computer or hypercomputer [1]. Hypercomputing is an interesting and economical alternative to super computers for memory bound applications. Some research activities have tried to exploit the computing power of hypercomputers, such as PVM or Linda. PVM offers a message passing programming model. Its inconveniences are ....

C.H. Cap and V. Strumpen. Efficient parallel computing in distributed workstation environments. Parallel Computing, 19:1221--1234, 1993.

No context found.

C. H. Cap and V. Strumpen, `Efficient parallel computing in distributed workstation environments', Parallel Computing, 19, (11), 1221--1234 (1993).

.... at the easy disposal of a single application, software platforms were developed, which use the performance of the individual workstations to form a (virtual) parallel computer or hypercomputer , allowing to program a network of workstations just like a distributed memory parallel computer [9] [10]. With this technique a performance or more than 1 Gflops [16] and 20 Gips [8] can be reached, exploiting the computing power of otherwise idling workstations. 30] estimates that 90 of Lawrence Livermore National Laboratory s workload could be solved with this technology. There is already a ....

....[16] and 20 Gips [8] can be reached, exploiting the computing power of otherwise idling workstations. 30] estimates that 90 of Lawrence Livermore National Laboratory s workload could be solved with this technology. There is already a large number of research prototypes like PVM [15] Parform [10], Picl [14] P4 [6] Parmacs [17] Lips [25] or POSYBL [27] and there are even commercialised solutions like LINDA [1] or Express [23] Examples demonstrate, that for special applications the hypercomputer is a cheaper [22] and faster [8] choice than a traditional supercomputer, and international ....

[Article contains additional citation context not shown here]

C. H. Cap and V. Strumpen, Efficient Parallel Computing in Distributed Workstation Environments. Parallel Computing, 19(11), 1993.

No context found.

Clemens H. Cap and Volker Strumpen. Efficient Parallel Computing in Distributed Workstation Environments. Parallel Computing, 19:1221--1234, 1993.

No context found.

Clemens H. Cap, Volker Strumpen, Efficient Parallel Computing in Distributed Workstation Environments, Parallel Computing, 19(1993), pp. 1221-1234

No context found.

Clemens H. Cap and Volker Strumpen. Efficient Parallel Computing in Distributed Workstation Environments. Parallel Computing, pages 1221--1234, 1993.

No context found.

C. H. Cap and V. Strumpen, `Efficient parallel computing in distributed workstation environments', Parallel Computing, 19, 1221--1234 (1993).

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