A. Mahanti and C. J. Daniel, A SIMD approach to parallel heuristic search, Arti cial Intelligence 60 (1993) 243-282.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....it contains, and the worker s own cluster is not excluded. When subproblems are forced to be released through the startScatter mechanism, they are always scattered to a random cluster. To supplement scattering, PICO also uses a form of rendezvous load balancing that resembles CMMIP s [8,10] [23] and [18] also contain earlier, synchronous applications of the same basic idea. This procedure also has the important side e ect of gathering and distributing global information on the amount of work in the system, which in turn facilitates control of the scattering process, and is also critical ....

A. Mahanti and C. J. Daniel, A SIMD approach to parallel heuristic search, Arti cial Intelligence 60 (1993) 243-282.

.... parallel SIMD (single instruction multiple data) machines are inherently unsuitable for parallel heuristic search (Bobrow, 1993) several researchers have implemented heuristic search routines (IDA ) on SIMD architectures with impressive results (Cook Lyons, 1993; Powley, Ferguson Korf, 1993; Mahanti Daniels, 1993). In this work, portions of the search tree are given to the processors, each of which performs a heuristic search. Previous work has also dealt with search on MIMD machines; for example, Rao and Kumar discuss parallel depth first search (Kumar Rao, 1987; Rao Kumar, 1987) In this mold, Cook ....

Mahanti, A. & Daniels, C. (1993). A SIMD approach to parallel heuristic search. Artificial Intelligence, 60, 243--282.

....is controlled by the apparent workload of the cluster relative to the entire system, and the parameters minNonLocalScatterProb, targetNonLocalScatterProb, and maxNonLocalScatterProb. To supplement scattering, PICO also uses a form of rendezvous load balancing that resembles CMMIP s [8, 10] [22] and [17] also contain earlier, synchronous applications of the same basic idea. This procedure also has the important side effect of gathering and distributing global information on the amount of work in the system, which in turn facilitates control of the scattering process, and is also critical ....

A. Mahanti and C. J. Daniel, A SIMD approach to parallel heuristic search, Artificial Intelligence 60 (1993) 243-282.

....have autonomy at the level of SPMD architectures (single program multiple data streams) Several authors have shown that SIMD systems are indeed capable of doing data parallel combinatorial search. Iterative deepening search for solutions to the well known 15 puzzle on an SIMD system was studied [32], 33] where the primary focus was on load balancing. A more theoretical study of the load balancing problem for SIMD combinatorial search can be found in [34] in which optimal conditions are derived for triggering load balancing phases of the computation; they also test their method using the ....

A. Mahanti and C.J. Daniels, "A SIMD Approach to Parallel Heuristic Search," Artificial Intelligence, vol. 60, pp. 243-282, 1993.

....many approaches have already been studied [4] Recently, parallel B B has been investigated for SIMD computers. A global approach is to map busy to idle PEs in a one to one correspondence and to transfer nodes by global communication. This approach has been applied to the IDA algorithm in [6, 7]. Static or dynamic trigger conditions indicate when load balancing is reasonable (see [3] also) This research work was founded by the Deutsche Forschungsgemeinschaft (DFG) with a stipend in the frame of a Graduiertenkolleg of the Computer Science Department, University of Karlsruhe. The work ....

....because when one process is triggered all PEs are forced to execute this process. In the B B algorithm, after each iteration, a trigger mechanism tells whether load balancing is necessary. In former research, two types of mechanisms for triggering synchronous load balancing have been used (see [3, 6, 7]) First, the static trigger S x indicates load balancing if the number of active PEs A decreases a fixed fraction x of all available PEs p. Thus, A x P. Second, the dynamic trigger D P or D K adapt themselves and indicate load balancing if the gain by balanced load is greater than the ....

Mahanti A., Daniels C. J., 1993, "A SIMD approach to parallel heuristic search", Artificial Intelligence, vol 60, no 2, pp 243-282.

....was only targeted for MIMD computers and it was of common sense that dynamic load balancing was not useful for SIMD parallel machines. Powley and al [1,2] show that dynamic load balancing strategies could be applied with success on SIMD computers. Other researchers improved this work (see [3 5]) We proposed several load balancing schemes especially suited for data parallel algorithms on SIMD computers [6] This paper presents the theoretical analysis of several dynamic load balancing strategies and the different results we obtained on our 16k processors MasPar MP 1. The paper is ....

Ambuj Mahanti and Charles J. Daniels. A SIMD approach to parallel heuristic search. Artificial Intelligence, 60:243, 1993.

....processors. The algorithm of Evett et al. 1990] performs a mapping of the search space onto the processing elements of a SIMD machine. This allows to eliminate duplicate states at the cost of an increased communication overhead. Two other approaches, SIDA by Powley et al. 1993] and IDPS by Mahanti and Daniels [1993] also run on the CM 2. From these, SIDA is probably the fastest parallel IDA implementation, solving all 100 problem instances [Korf, 1985] of the 15 puzzle in 2.245 hours. # # # # # #c c c c c c oe root node oe t init oe tbase base level for load balancing . # # # c c c n i # # # c ....

A. Mahanti, C.J. Daniels. A SIMD approach to parallel heuristic search. Art. Intell. 60(1993), 243 -- 282.

....all processors run out of work and the next iteration is started. This continues until one of the processors discovers the goal node, and the search is subsequently terminated. PIDA achieves an average speedup of 28 on 30 processors of a Sequent Balance 21000. IDPS by Mahanti and Daniels [58] and SIDA by Powley, Ferguson and Korf [76] are similar algorithms for handling parallel IDA search on SIMD machines. Both algorithms partition a set of frontier nodes to each processing element. Each processing element uses depth first search independently. When a number of processors go idle, ....

A. Mahanti and C. Daniels. A SIMD Approach to Parallel Heuristic Search. Artificial Intelligence, 60:243--282, 1993. (66)

....there has been great deal of research into serial heuristic search, and some into coarsegrained MIMD search 7 [7, 10, 23] research into SIMD search algorithms is just beginning. Other than PRA , the only attempts we know of to develop SIMD search algorithms have been parallelizations of IDA [20, 22, 11, 12]. The earliest attempt at a SIMD version IDA was done by Powley and Korf [20] They parallelized IDA by allowing individual processors to execute separate depth first searches to different thresholds (cost bounds) The first solution found is not guaranteed to be optimal, so search continues ....

Mahanti, A. and Daniels, C.J. "A SIMD Approach to Parallel Heuristic Search," Workshop on Parallel Computing of Discrete Optimization Problems, Univ. Minnesota, May 22-24, 1991.

No context found.

A. Mahanti and C.J. Daniels, A SIMD Approach to Parallel Heuristic Search, Artificial Intelligence, 10 (1993), pp. 243--282.

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