Lee, P.-Z. and Chen, W. Y. 1997. Generating global name-space communication sets for array assignment statements. Technical Report TR-IIS-97-016 (Oct.), Institute of Information Science, Academia Sinica, Taipei, Taiwan, on available via WWW http://www.iis.sinica.edu.tw/#leepe/PAPER/tr97016.ps. Home/Search Document Details and Download
Summary Related Articles Check |
This paper is cited in the following contexts:
Automatic Data and Computation Decomposition on Distributed.. - Lee, Kedem (1999) (1 citation) (Correct)
....data distributions and scheduling computation based on the owner computes rule [Gupta and Banerjee 1992; Lee 1995b; Li and Chen 1991b] became prominent. Data redistribution between program fragments can also be optimized by comparing the relative costs of di#erent data distribution schemes [Lee 1997]. For other approaches see [Kennedy and Kremer 1998; Kremer 1995; Palermo 1996] For a complete survey of other data distribution techniques see [Lee 1997] In general, component alignment approaches are very promising for DMPCs, because dimensions on each data array can be distributed ....
.... Data redistribution between program fragments can also be optimized by comparing the relative costs of di#erent data distribution schemes [Lee 1997] For other approaches see [Kennedy and Kremer 1998; Kremer 1995; Palermo 1996] For a complete survey of other data distribution techniques see [Lee 1997]. In general, component alignment approaches are very promising for DMPCs, because dimensions on each data array can be distributed independently among one another, and following the HPF standard. What needs to be done is the combination of determining data distributions for data spaces and ....
[Article contains additional citation context not shown here]
Lee, P.-Z. and Chen, W. Y. 1997. Generating global name-space communication sets for array assignment statements. Technical Report TR-IIS-97-016 (Oct.), Institute of Information Science, Academia Sinica, Taipei, Taiwan, on available via WWW http://www.iis.sinica.edu.tw/#leepe/PAPER/tr97016.ps.
Automatic Data and Computation Decomposition on Distributed.. - Lee, Kedem (2002) (1 citation) (Correct)
....(2) the computational load among the PEs is balanced, and (3) communication is minimized. Currently, stride alignment constraints can be satisfied, but we still depend on table look up heuristics to select suitable block sizes which compromise both load balance and communication overhead [27]. Although for a specific class of problem, block sizes can be determined based on finding the optimal tile sizes as described in Section 6:3, finding optimal block sizes for general cases is still an open issue. Explanation: Block sizes have to satisfy stride alignment constraints, otherwise, ....
....is required. For example, if A(l 1 is 1 ) is axis aligned with C(l 2 is 2 ) A is distributed by cyclic(b 1 ) C is distributed by cyclic(b 2 ) and b 1 =s 1 = b 2 =s 2 ; then their stride alignments are matched. For details of generating communication sets, interested readers can refer to [27]. Next, if the iteration space is not rectangular, in order to maintain load balance, small block sizes are preferred. For example, if the iteration space is a pyramid (such as the iteration space of an LU decomposition) or a triangle (such as the iteration space of a triangular linear system) ....
P.-Z. Lee and W. Y. Chen. Generating global name-space communication sets for array assignment statements. Technical Report TR-IIS-97-016, Institute of Information Science, Academia Sinica, Taipei, Taiwan, October 1997, on available via WWW http://www.iis.sinica.edu.tw/žleepe/PAPER/tr97016.ps.
Efficient Algorithms for Data Distribution on Distributed Memory.. - Lee (1997) (11 citations) (Correct)
....adopt Li and Chen s heuristic algorithm by regarding our directed component affinity graphs as being undirected. The directions of edges, which indicate parent child relations, can be used to determine block sizes of data distribution so that communication sets can be represented by closed forms [29]. The directions of edges also are used in a code generation phase and will be used to determine the direction of the data communication according to the owner computes rule. For completeness, in Fig. 2, we present a very brief version of the component alignment algorithm; however, interested ....
P.-Z. Lee and W. Y. Chen. Generating global name-space communication sets for doall statements. Submitted for publication, on available via WWW http://www.iis.sinica.edu.tw/žleepe/PAPER/comm97.ps.
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