E.T. Kalns and L.M. Ni, "DaReL: A Portable Data Redistribution Library for Distributed-Memory Machines," Proc. 1994 Scalable Parallel Libraries Conf. II, Oct. 1994.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....formal processors onto actual processors such that the global communication can be translated to the local communication in a certain processor group. Since the communication is local to a processor group, one can reduce communication conflicts when performing a redistribution. Kalns and Ni [10] [11] proposed a processor mapping technique to minimize the amount of data exchange for BLOCK to BLOCK CYCLIC(c) redistribution and vice versa. Using the data to logical processors mapping, they showed that the technique can achieve the maximum ratio between data retained locally and the total amount ....

E.T. Kalns and L.M. Ni, "DaReL: A Portable Data Redistribution Library for Distributed-Memory Machines," Proc. 1994 Scalable Parallel Libraries Conf. II, Oct. 1994.

....to renumbering the destination process numbers, and although this is permitted within the HPF specification it would not be directly compatible with the form of block cyclic data distribution assumed in ScaLAPACK. Kalns and Ni have used this approach in their redistribution library Darel [6] which is implemented using MPI F, the custom implementation of MPI for the IBM SP 1 and SP 2. Thakur et al. 11] have implemented an algorithm on the Intel Paragon for transforming between two block cyclic data distributions by examining each index of the array to determine its destination ....

E. T. Kalns and L. M. Ni. Darel: A portable data redistribution library for distributed memory machines. In Proceedings of the 1994 Scalable Parallel Libraries Conference, October 1994. Redistribution of Block-Cyclic Data Distributions Using MPI 27

....on the efficiency of computing send receive processor sets rather than on the efficiency of the actual data exchange portion of redistribution. We have found that the latter operation can be several orders of magnitude more costly, in terms of execution time than send and receive set determination [15]. Data exchange on distributed memory machines can be performed with either point to point or collective communication message passing primitives. Techniques for efficient multiple scatter and gather operations are presented in [16] McKinley et al. 17] survey issues regarding all to all ....

....the compiler performs a table lookup to determine neighboring lpid relationships. The extension to boundary processors is straightforward and is excluded from the present discussion. 5. 3 Performance We have integrated the optimal mapping technique into a data redistribution library, DaReL [15], which is written in C using MPI F [22] The performance results were obtained on the IBM SP x at Argonne National Laboratory [23] To assess the run time performance of the optimal mapping technique, we compare data redistribution execution times using the optimal technique with redistribution ....

[Article contains additional citation context not shown here]

E. T. Kalns and L. M. Ni, "DaReL: A portable data redistribution library for distributedmemory machines," in Proceedings of the 1994 Scalable Parallel Libraries Conference II, Oct. 1994.

No context found.

E. T. Kalns and L. M. Ni. DaReL: A portable data redistribution library for distributedmemory machines. In Proceedings of the 1994 Scalable Parallel Libraries Conference II, October 1994.

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