R. J. Anderson and J. C. Setubal. On the parallel implementation of Goldberg's maximum flow algorithm. In Proceedings of the 4th annual Symposium on Parallel Algorithms and Architectures, 1992.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....The performance data for running our program on a SPARC 10 41 workstation is shown in Table 8.2. 8. 2 Parallel Implementation We implemented the simple PRAM algorithm for finding a smallest 2edge connectivity augmentation given in [ET76, Sor88] on a massively parallel SIMD computer MasPar MP 1 [Mas92d] An introduction of the MasPar is given in Part II of this thesis. This algorithm runs in O(log n) time on a CRCW PRAM using a linear number of processors given the adjacency list of the input graph, where n is the number of vertices in the input graph. The 184 c w test 1 test 2 test 3 test ....

....1,975 1,975 1.08 1.17 1.10 1.13 1.12 Table 8.2: Performance data for finding a smallest biconnectivity augmentation on a SPARC 10 41. MasPar computer that we used had 16,384 processors where each processor is about 230 times slower than a SUN SPARC 10 41. We used the parallel language MPL [Mas92b, Mas92c] that is an extension of the C programming language [KR88] An introduction of the MPL is also given in Part II of this thesis. Since the MPL does not support the use of virtual processors, we implemented the algorithm only to handle the case when the input size is no more than the ....

[Article contains additional citation context not shown here]

R. Anderson and J. Setubal. On the parallel implementation of Goldberg's maximum flow algorithm. In Proc. 4th ACM Symp. on Parallel Algorithms and Architectures, pages 168--177, 1992.

....we implemented several different parallel algorithms for the connected components problem, including one randomized algorithm, and tested our code with respect to various fine tuning techniques. Related work on implementing combinatorial algorithms on massively parallel machines can be found in [1, 3, 4, 8, 9, 10, 11, 15, 16, 18, 19, 30, 38, 39, 41, 48]. Also there has been work reported on implementing combinatorial algorithms on a vector super computer [16, 45, 49] and on a distributed memory machine [29] The rest of the paper is organized as follows. Section 2 describes the algorithms implemented which includes an algorithm that we devised ....

R. Anderson and J. Setubal, On the parallel implementation of Goldberg's maximum flow algorithm, Proc. 4th ACM Symp. on Parallel Algorithms and Architectures, 1992, pp. 168-- 177.

....by an IBM graduate fellowship. 1 Introduction This paper describes an on going project for implementing parallel graph algorithms on the massively parallel machine MasPar MP 1. There has been a fair amount of prior work on implementing parallel algorithms on massively parallel machines [1, 5, 9, 10, 11, 25, 29] since the completion of the first phase of our project reported in [13] However, most of this work has been targeted towards solving problems that are highly structured and are not very difficult to scale up. The focus of our work is on solving graph theoretical problems for which the algorithms ....

R. Anderson and J. Setubal. On the parallel implementation of Goldberg's maximum flow algorithm. In Proc. 4th ACM Symp. on Parallel Algorithms and Architectures, pages 168--177, 1992.

....to RCupd) and invalidate (compared to RCinv) traffic (see Table 5) in RCexp also tends to bring down the mutex stall time. 6.4 Maxflow The Maxflow application finds the maximum flow from the distinguished source to the sink, in a directed graph with edge capacities. In the implementation [3], each processor accesses a local work queue for tasks to perform. These may in turn generate new tasks which are added to this local work queue. Each task involves read and write accesses to shared data. The local queues of all processors interact via a global queue for load balancing. Thus the ....

R. J. Anderson and J. C. Setubal. On the parallel implementation of goldberg's maximum flow algor ithm. In 4th Annual ACM Symposium on Parallel Algorithms and Archite ctures, pages 168--77, June 1992.

....sizes) a mesh interconnect with a link latency of 1.6 CPU cycles per byte; store buffer of size 4 entries; and a merge buffer of one cache block. The applications we studied include Cholesky and Barnes Hut from SPLASH suite [12] Integer Sort from the NAS parallel benchmark suite [3] and Maxflow [2]. Cholesky performs a factorization of a sparse positive definite matrix. The sparse nature of the matrix results in an algorithm with a data dependent access pattern. Sets of columns having similar non zero structure are combined into supernodes. Supernodes are added to a central work queue if a ....

....Read stall Execution time (10000 cycles) Figure 3: IS is well defined statically. The problem size considered is 32K integers with 1K buckets. The Maxflow application finds the maximum flow from the distinguished source to the sink, in a directed graph with edge capacities. In the implementation [2], each processor accesses a local work queue for tasks to perform. These may in turn generate new tasks which are added to this local work queue. Each task involves read and write accesses to shared data. The local queues of all processors interact via a global queue for load balancing. Thus the ....

R. J. Anderson and J. C. Setubal. On the parallel implementation of goldberg's maximum flow algorithm. In 4th Annual ACM Symposium on Parallel Algorithms and Architectures, pages 168--77, June 1992.

.... pppppppppppppppppppp pppppppppppppppppppp pppppppppppppppppppp pppppppppppppppppppp pppppppppppppppppppp pppppppppppppppppppp pppppppppppppppppppp pppppppppppppppppppp Read stall 10000 cycles Figure 2: IS suite [9] Integer Sort from the NAS parallel benchmark suite [2] and Maxflow [1]. As we mentioned earlier (see Section 2) by definition the z machine will not have any write stall or buffer flush times. The inherent communication cost in the application will manifest as read stall times due to the timing relationships for the memory accesses from the producers and consumers ....

R. J. Anderson and J. C. Setubal. On the parallel implementation of goldberg's maximum flow algorithm. In 4th Annual ACM Symposium on Parallel Algorithms and Architectures, pages 168--77, June 1992.

....abmp algorithm. The input classes considered are three variations of random bipartite graphs; they were designed to be representative of bipartite matching problems that might arise in practice. The largest instances solved have 120000 vertices. On another set of experiments, Anderson and Setubal [AS92b] have shown how to obtain good speed ups for Goldberg s maximum flow algorithm when implemented on a shared memory multiprocessor with a small number of processors. So it is again natural to ask whether a parallel implementation of Goldberg s algorithm specialized for unweighted bipartite graphs ....

....matching is the greedy one, in which a match is attempted between each vertex in U and the first vertex on its adjacency list. For the input classes described in this paper, this initial greedy matching matches over 80 of the vertices. From experience gained with the maximum flow problem [AS92b] it was feared that not much parallelism would be available to match the remaining 20 or so vertices. Yet in this paper we report success in obtaining speed ups for all inputs. The numbers obtained look somewhat modest, ranging from 2.4 to 3.2 with 12 processors on the largest instances, but the ....

[Article contains additional citation context not shown here]

R. J. Anderson and J. C. Setubal. On the parallel implementation of Goldberg's maximum flow algorithm. In Proc. 4th Symp. on Parallel Algorithms and Architectures, pages 168--177. ACM Press, 1992.

No context found.

R. J. Anderson and J. C. Setubal. On the parallel implementation of Goldberg's maximum flow algorithm. In Proceedings of the 4th annual Symposium on Parallel Algorithms and Architectures, 1992.

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