P. Boulet, J. Dongarra, Yves Robert, and Frederic Vivien. Tiling for heterogeneous computing platforms. Technical Report UT-CS-97-373, University of Tennessee, Knoxville, 1997.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....There are p available processors, numbered from 1 to p, which are assigned columns of tiles. When targeting a homogeneous NOW, a natural way to allocate tile columns to physical processors using a pure cyclic allocation [15, 14, 1] For heterogeneous NOWs, we use a re ned periodic allocation (see [6]) which proves e cient both theoretically and experimentally: we have run several MPI experiments. We point out that purely cyclic allocations do not perform well, while they would be the outcome of a greedy master slave strategy. Indeed, processors will be allocated the rst p columns in any ....

....solution consists in allocating panels of B tile columns to the p processors in a periodic fashion. Inside each panel, processors receive an amount of columns inversely proportional to their speed. Within each panel, the work is weell balanced, and dependences do not slow down the execution. See [6] for details. Finite di erence stencil computations In this example we study a tiled version of the FermiPasta Ulam one dimensional relaxation problem [18] 00000000000000000000000001111111111111111111111111 Spring Molecule x1 x2 x l m a = P F m x i = k(x i 1 x i ) K(x i 1 x i ) k(x ....

P. Boulet, J. Dongarra, Yves Robert, and Frederic Vivien. Tiling for heterogeneous computing platforms. Technical Report UT-CS-97-373, University of Tennessee, Knoxville, 1997. To appear in the journal Parallel Computing.

No context found.

P. Boulet, J. Dongarra, Yves Robert, and Frederic Vivien. Tiling for heterogeneous computing platforms. Technical Report UT-CS-97-373, University of Tennessee, Knoxville, 1997.

No context found.

P. Boulet, J. Dongarra, Yves Robert, and Frederic Vivien. Tiling for heterogeneous computing platforms. Technical Report UT-CS-97-373, University of Tennessee, Knoxville, 1997. 23

No context found.

P. Boulet, J. Dongarra, Yves Robert, and Fr#d#ric Vivien. Tiling for heterogeneous computing platforms. Technical Report UT-CS-97-373, University of Tennessee, Knoxville, 1997.

....) T ) t proc(T ) T com (T 0 ) otherwise To increase the granularity and to minimize the number of communications, columns of tiles (rather than single tiles) are allocated to processors. So a computational chunk will be a tile column. Columnwise allocations are asymptotically optimal [5]. When targeting a homogeneous NOW, a natural way to allocate tile columns to physical processors is to use a pure cyclic allocation [16, 14, 2] in HPF words, this is a CYCLIC(1) distribution of tile columns to processors) For heterogeneous NOWs, we use a periodic allocation based upon the ....

P. Boulet, J. Dongarra, Yves Robert, and Frederic Vivien. Tiling for heterogeneous computing platforms. Technical Report UT-CS-97-373, University of Tennessee, Knoxville, 1997.

....) T ) t proc(T ) T com (T 0 ) otherwise To increase the granularity and to minimize the number of communications, columns of tiles (rather than single tiles) are allocated to processors. So a computational chunk will be a tile column. Columnwise allocations are asymptotically optimal [4]. When targeting a homogeneous NOW, a natural way to allocate tile columns to physical processors is to use a pure cyclic allocation [15, 13, 2] in HPF words, this is a CYCLIC(1) distribution of tile columns to processors) For heterogeneous NOWs, we use a periodic allocation based upon the ideas ....

P. Boulet, J. Dongarra, Yves Robert, and Frederic Vivien. Tiling for heterogeneous computing platforms. Technical Report UT-CS-97-373, University of Tennessee, Knoxville, 1997.

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