Cap, C. H. Massive Parallelism with Workstation Clusters - Challenge or Nonsense? Technical Report IFI-TR 94.01. Department of Computer Science. CH-8057 Zurich, Switzerland.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....a WAN environment, too. The reason for this constraint stems from the fact that firstly, we can experience significant communication problems between distant computers, and secondly it is not likely, at this stage, that we can use resources that are not within our institution. As reported by Cap [Cap 93] workstation cluster computing is limited by the communication bottleneck to small and medium sized configurations of up to 30 workstations for most applications. 5 The author concludes that for special applications and algorithms massive parallelism is possible in a wide area network and ....

Cap, C. H. Massive Parallelism with Workstation Clusters - Challenge or Nonsense? Technical Report IFI-TR 94.01. Department of Computer Science. CH-8057 Zurich, Switzerland.

.... [1] 2] Regarding the price performance ratio, workstation clusters (WSC) can achieve better results compared to sequential or parallel supercomputers [3] 4] An example of the possibilities of WSC was presented through the execution of a parallel application simultaneously on more than 800 WS [5]. Communication between processing elements during the execution of a parallel application is a main bottleneck for all distributed memory systems and especially for clusters. Applications is connected to n different network channels. In the example of a two dimensional CNA cluster (Fig. 1) the ....

C. H. Cap: "Massive Parallelism with Workstation Clusters - Challenge or Nonsense", Proceedings of the High Performance Computing and Networking Conference, HPCN 94, Munich, 1994

....and their current load, executive processes for setting up the application tasks with the corresponding communication links, and load sensor daemons for determining the CPU work load by running small benchmark routines at regular intervals. An extension of the Parform project, named PARC [10, 38], emphasizes on the efficient management of distributed applications in very large WAN clusters. Based on a tree shaped logical interconnection topology,thePARC system provides a programming interface and tools for the administration of large WAN metacomputers with several hundred nodes. HiCon. ....

....program s communication is not affected by the monitoring. Expected Bandwidths and Latencies. Ethernet LANs can connect as muchas255 workstations, but an average LAN comprises about 20 machines. In practice, more than 50 of the 10 Mbps physical bandwidth can be expected in a typical run [10] of a parallel PVM application. However, a utilization of more than 20 is already considered as a high load. Ethernet s CSMA CD protocol will then suffer many collisions, thereby further reducing the maximum attainable throughput. On a WAN, the attainable bandwidths are muchlower. As an example, ....

C.H. Cap. Massive parallelism with workstation clusters -- Challenge or nonsense? Procs. HPCN-94, Munich, Springer LNCS 797 (1994), 42--52.

....that are used within distributed environments are much more concerned with the cost of communication than applications executing on massively parallel environments. This is because each processing unit must communicate with another over a network which has a higher cost of communication [CAP93]. Some distributed environments s are Berkeley s NOW [CULL97] MOSIX [BARA98] Amoeba [TANE91] Sprite [OUST88] V [CHER88] Chorus [ROZI92] Globe [STEE97] Beowulf [RIDG97] and RHODOS [DePA95] The type of parallel execution that needs to be performed will most certainly play an important part ....

C.H. Cap. "Massive Parallelism with Workstation Clusters - Challenge or Nonsense?", Technical Report IFI-TR 94.01, Department of Computer Science, University of Zurich, Zurich, Switzerland, December 14, 1993.

....suitable applications, which have large granularity, a high degree of parallelism and low sensitivity to communications latency, such metacomputers can form a practical, viable and scientifically useful resource. One feasibility study used over 800 workstations distributed across five continents [12]. This wide variety of solutions is described under a range of names: distributed computing; cluster computing; metacomputing etc. 5.1 Adapting Existing Environments for Java The majority of HPC applications on distributed memory systems are implemented using message passing programming using ....

C.H. Cap, "Massive Parallelism with Workstation Clusters -- Challenge or Nonsense" in Proceedings of the High Performance Computing and Networking Conference HPCN `94 Europe, Munich 1994, Springer LNCS 797, 42--52, 1994, http://www.tec.informatik.uni-rostock.de/IuK/ publications/index.html.

....a superposition without a mixing up of independant resources. This results in a set of orthogonal resource fields. Within such a resource field a resource is located with a gradient search. Keywords: network computing, resource propagation, implicit load balancing 1. INTRODUCTION Several works [8, 9] have shown that workstations run idle most of the time. The objective of the Hypercomputer Project [2] is, to create a virtual computer called Hypercomputer, which uses idle times of workstations connected via the internet to run programs. As a result of the experiences made with the former ....

Clemens H. Cap, Massive Parallelism With Workstation Clusters -- Callange or Nonsense, Proceedings of the High Performance Computing and Networking Conference, HPCN'94, Munich, 1994

....as they have on super computers. They have also calculated that on average, they only actually use 15 of the capacity of these workstations, but that this small amount still accounts for five times the computational power of all of their super computers [23] Research at the University of Zurich [25] made use of over 800 individual workstations to show that it is even feasible to make use of several Local Area Networks connected by a Wide Area Network for some classes of problems. Obviously it would be of enormous benefit if it were possible to make use of the remaining 85 , and it is this ....

C. H. Cap. Massive parallelism with workstation clusters --- challenge or nonsense? Technical Report IFI-TR 94.01, Department of Computer Science, University of Zurich, December 1993.

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