K. Rokusawa, N. Ichiyoshi, T. Chikayama, and H. Nakashima. An efficient termination detection and abortion algorithm for distributed processing systems. In International Conference on Parallel Processing, Vol. I, pp. 18--22, 1988. also published as ICOT-TR 341.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....these two programs are identical to the original one in Figure 3.1. Some concurrent applications, such as search problems, invoke a large number of threads, where termination detection of all the threads is a difficult problem because there is no global control. Several algorithms (cf. [85, 104]) have been proposed to solve this problem. However, when we incorporate a termination detection algorithm into a na ively written concurrent program, we often have to modify the structure of the original program, such as additional parameters to each function definition and invocation, sending ....

....and eval fork, respectively. At the frontmost layer, an evaluator class is defined for each termination detection algorithm. Figure 3.20 presents the simplest one in which an acknowledgment message is returned for each fork invocation. Other algorithms e.g. the one using global weight[85, 104] can be implemented in similar ways. The overview of the algorithm is as follows (operations written in the slanted font are performed at the meta level) 1) A method function is invoked. 2) A reply box is created for each child. 3) A child (i.e. sub computation) is forked. The reply box ....

[Article contains additional citation context not shown here]

K. Rokusawa, N. Ichiyoshi, T. Chikayama, and H. Nakashima. An efficient termination detection and abortion algorithm for distributed processing systems. In International Conference on Parallel Processing, Vol. I, pp. 18--22, 1988. also published as ICOT-TR 341.

....or and parent and a tree structure is built on this relationship. A parent PE cannot be passive until all of its children are passive. When the root of the tree becomes passive, termination is detected [7] Another class of algorithms is based on the credit recovery or weight throwing principle [20, 22, 25, 26]. 3 RING ALGORITHMS In this section we describe termination detection algorithms for the AP and APS models when processors use a token passing logical ring network on processors PE 0 ; PEN Gamma1 . Observe that only control messages use the ring structure and that communication ....

K. Rokusawa, N. Iciyoshi, T. Chikayama, and H. Nakashima. An efficient termination detection and abortion algorithm for distributed processing systems. In Proceedings of the International Conference on Parallel Processing, pages 18--22, 1988.

....are identical to the original one in Figure 2. 3.3 Termination Detection Some parallel applications, such as search problems, invokes a large number of threads, where termination detection of all the threads is a difficult problem because there is no global control. Several algorithms (cf. [9, 14]) have been proposed to solve this problem. However, when we incorporate a termination detection algorithm into a naively written parallel program, we often have to modify the structure of the original program, such as adding parameters to each function definition and invocation, sending control ....

....and eval fork, respectively. At the leftmost layer, an evaluator class is defined for each termination detection algorithm. Here, we only present the simplest one in which an acknowledgment message is returned for each fork invocation. Other algorithms e.g. the one using global weight[9, 14] can be implemented in similar ways. The overview of the algorithm is as follows (operations written in the slanted font are performed at the meta level) 1) A method function is invoked. 2) A reply box is created for each child. 3) A child (sub computation) is forked. The reply box is passed ....

K. Rokusawa, N. Ichiyoshi, T. Chikayama, and H. Nakashima. An efficient termination detection and abortion algorithm for distributed processing systems. In International Conference on Parallel Processing, volume I, pages 18--22, 1988. also published as ICOT-TR 341.

....method [22] We evaluate their respective performances and compare their advantages disadvantages. Termination of marking must also be detected with low message traffic, but it is difficult to do so efficiently in distributed environments; our algorithm uses Weighted Throw Counting (WTC) scheme[17, 14] to reduce message traffic. We have implemented prototype version of our algorithm on Fujitsu AP1000, a multicomputer with 32 1024 nodes, and validated its correct operation and did several performance benchmarks, including the measurement of reduction in message traffic of global GC. Section 2 ....

....an acknowledgement message z per every mark message to indicate the arrival of the message, or the termination of marking from it. Thus, the number of messages required for termination detection would be more than twice the number of mark messages. In order to reduce this cost, we use WTC scheme[17, 14] to detect quiescent state of marking. Mark messages and marking local collectors are given positive weights by the GC host. The GC host finds the quiescent state of marking when all weights are returned to the GC host. The GC host gives positive weights to all the local collectors at the ....

K. Rokusawa, N. Ichiyoshi, T. Chikayama, and H. Nakashima. An Efficient Termination Detection and Abortion Algorithm for Distributed Processing Systems. In Proc. of ICPP'88, volume 1, pages 18--22, 1988.

.... in Hand craft and Meta are: 1) Locality control child nodes (objects) at deep levels in the search tree are created at the same processor as their parents in order to reduce remote communication overhead (the default is to randomly choose a processor) 2) Weighted termination detection[19] weight is propagated along the search tree in order to detect the end of a search process. By default, the detection is achieved by collecting acknowledgments in the search tree; therefore, intermediate search nodes cannot be released until all its descendant nodes terminate. The meta level ....

K. Rokusawa, N. Ichiyoshi, T. Chikayama, and H. Nakashima. An efficient termination detection and abortion algorithm for distributed processing systems. In International Conference on Parallel Processing, vol. I, pp. 18--22, 1988. also published as ICOT-TR 341.

....as a problem because user process failure could migrate up to the operating system in a naive implementation. ICOT developed protected tasks called shoen [40] similar to work by I. Foster [24] Functional mechanisms for shoen management have been implemented at ICOT over the past four years [58, 63, 64], although empirical measurements of the operating system running applications programs have not yet been analyzed. Intra cluster mechanisms include termination and deadlock detection [43] and resource caching; inter cluster mechanisms include weighted throw counts. Without further empirical ....

K. Rokusawa, N. Ichiyoshi, T. Chikayama, and H. Nakashima. An Efficient Termination Detection and Abortion Algorithm for Distributed Processing Systems. In International Conference on Parallel Processing, pages 18--22, Penn State, August 1988.

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