Part of this work has been presented in [17]. This paper analyzes the performances of portable distributed priority queues by examining the theoretical features required and by comparing various implementations. In spite of intrinsic bottlenecks and induced hot-spots, we argue that tree topologies are attractive to manage the natural centralized control required for the deletemin operation in order to detect the site which holds the item with the largest priority. We introduce an original perfect balancing to cope with the load variation due to the priority queue operations which continuously modify the overall number of items in the network. For comparison, we introduce the d-heap and the binomial distributed priority queue. The purpose of this experiment is to convey, through executions on Cray-T3D and MeikoT800, an understanding of the nature of the distributed priority queues, the range of their concurrency and a comparison of their efficiency to reduce requests latency. In particu...

) Ajay K. Gupta ? Andreas G. Photiou Western Michigan University Lake States Insurance Company Kalamazoo, MI 49008, USA Traverse City, MI 49685, USA Abstract. We consider efficient algorithms for priority queues on distributed memory multiprocessors, such as nCUBE, iPSc, MPP and looselycoupled systems consisting of networked workstations. For a p-processor distributed memory multicomputer P and n data items in the priority queue, n ? p, we investigate two priority queues; horizontally sliced and vertically sliced. Both of these achieve load balance, i.e. at most \Theta(n=p) data items are stored at every processor of P . Horizontally sliced priority queue allows deletions and insertions of \Theta(p) items in time O( p bw øc + øpp log n) on hypercubic networks where øc is the communication time between a pair of processors, øp is the unit processing time and bw is the width of the communication channel between a pair of processors. Vertically sliced priority queue allows deletio...

. We analyze a class of adaptive integration algorithms on MIMD distributed memory systems. The integration region subdivided in the course of the adaptive process is the N-dimensional cube or simplex. At the subdivision of a subregion, the error behaves according to a prescribed model. The model is supported by the asymptotic behavior of the error for integrands which are continuously differentiable of a given order over all subregions with the possible exception of one subregion containing a vertex singularity of homogeneous type. Extensions are possible to deal with multiple vertex singularities. The analysis can also be applied to problems in other areas, as long as the task selection is based on a priority function which behaves according to a suitable model. Using an efficient management of the subregions, we show that a O(p= log p) speedup can be achieved on a p-processor hypercubic network, such as shuffle exchange, butterfly and hypercube. Furthermore, a speedup of O( p p) c...

We analyze a class of adaptive algorithms for integration over N-dimensional hyperrectangular or simplical regions, on distributed systems. An adaptive algorithm attempts to achieve the requested accuracy by refining the subdivision of the integration region, thus allowing for a concentration of subdivisions near singularities. At the subdivision of a region, the error behaves according to a prescribed model, relating the error of the parent region to that of its children. The analysis can also be applied to problems in other areas, as long as the task selection is based on a priority function which behaves according to a suitable model. Using an efficient management of the subregions, we show that a O(p= log p) speedup can be achieved on a p-processor hypercubic network, such as shuffle exchange, butterfly and hypercube. Furthermore, a speedup of O( p p) can be achieved on a p p \Theta p p mesh network. We also show that our algorithms compare favorably with well-known dynamic...

ectly, with this thesis. I would like to express my appreciation to my thesis committee. Dr. Ajay Gupta, who found time to help me out despite his many new responsibilities, Dr. Elise de Doncker, who put up with my apparent inability to understand anything that might have come from a math textbook, and Dr. Don Nelson, who took time out from his sabbatical to help with my thesis. In addition to helping me with my thesis, they have guided me through the beginning of the process of becoming a computer science researcher. I must thank Sue Moorian for her help on many administrative snafus, and Ken Thies and Orien Vandenbergh for their help in solving many system related problems that cropped up while working on ParInt these last few years. Without the help of all the members of the ParInt research group, past and present, the research environment that made this thesis possible would not have existed. And as always, credit is due to all of the scientists u

For maximum efficiency in a multiprocessor system the load should be shared evenly over all processors, that is, there should be no idle processors when tasks are available. The delay in a load sharing algorithm is the larger of the maximum time that any processor can be idle before a task is assigned to it, and the maximum time that it must wait to be relieved of an excess task. A simple parallel priority queue architecture for load sharing in a p-processor multiprocessor system is proposed. This architecture uses O(p log(n=p)) special-purpose processors (where n is the maximal size of the priority queue), an interconnection pattern of bounded degree, and achieves delay O(logp), which is optimal for any bounded degree system. 1 Introduction One advantage that multiprocessor computers have over uniprocessors is the ability to speed up computation by having the processors compute in parallel. The archetypal model studied is the PRAM, in which it is assumed that concurrent access to a ...

In this report, we propose new concurrent data structures and load balancing strategies for Branch-and-Bound (B&B)/A* algorithms in two models of parallel programming : shared and distributed memory. For the shared memory model (SMM), we present a general methodology which allows concurrent manipulations for most tree data structures, and show its usefulness for implementation on multiprocessors with global shared memory. Some priority queues which are suited for basic operations performed by B&B algorithms are described : the Skew-heaps, the funnels and the Splay-trees. We also detail a specific data structure, called treap and designed for A* algorithm. These data structures are implemented on a parallel machine with shared memory : KSR1. For the distributed memory model (DMM), we show that the use of partial cost in the B&B algorithms is not enough to balance nodes between the local queues. Thus, we introduce another notion of priority, called potentiality, between nodes that take...