Tyng--Ruey Chuang, Rong--Guey Chang, and Jenq Kuen Lee. Sampling and analytical techniques for data distribution of parallel sparse computation. In Eighth SIAM Conference on Parallel Processing for Scientific Computing. Minneapolis, Minnesota, USA, March 1997. 8 pages. SIAM Press.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....10 Figure 14. Two different ways to calculate the group by operation according to the lattice. no. of proc. LHS Tree of Lattice RHS Tree of Lattice 1 0.317 0.271 2 0.171 0.146 4 0.118 0.093 8 0.095 0.071 Table 2. Performance results of 3d group by. 4. Related Work Previous research work [1, 16, 3, 15] considered the distribution and compressed schemes related to one dimension and two dimension arrays. The work in [16] pioneered the research directions to annotate sparse notations and distributions for HPF style languages. In our work, we built an extension applicable to higherdimensional ....

Tyng--Ruey Chuang, Rong--Guey Chang, and Jenq Kuen Lee. Sampling and analytical techniques for data distribution of parallel sparse computation. In Eighth SIAM Conference on Parallel Processing for Scientific Computing. Minneapolis, Minnesota, USA, March 1997. 8 pages. SIAM Press.

....the performances of application programs that use sparse data sets. 1 Introduction The usage of array intrinsic functions in Fortran 90 provide a rich source of parallelism. In our recent work, we have been working on providing parallel sparse supports for array intrinsics of Fortran 90 [6, 4]. Our supporting library uses a two level design. In the low level routines, it requires the input sparse matrices to be specified with compression distribution schemes for array functions. In the high level representations, sparse array functions are overloaded with Fortran 90 array intrinsic ....

....research work on data alignments and distribution for dense arrays [5, 12, 13, 15] and from previous work on sparse array optimization [2] on four key elements. First, in our work, the selections of distributions schemes for sparse arrays are mainly based on the sparsity structures of arrays [6], while in previous work the selections of distributions schemes for dense arrays are mainly based on the index domains of arrays. Second, we need to select the compression schemes for sparse arrays, while there is no such need in the dense cases. The compression schemes being considered are ....

[Article contains additional citation context not shown here]

Tyng--Ruey Chuang, Rong--Guey Chang, and Jenq Kuen Lee. Sampling and analytical techniques for data distribution of parallel sparse computation. In Eighth SIAM Conference on Parallel Processing for Scientific Computing, March 1997.

....arguments (but not returning as results) dynamic storage allocation deallocation. Combining these language features with its module facility, one can develop very sophisticated user library in Fortran 90. We are in the process of building in Fortran 90 a library to support sparse array intrinsics [4]. By sparse array intrinsics, we mean that sparse arrays are used in the same ways as (dense) arrays in Fortran 90. That is, convenient notations can be used, elemental intrinsic functions can be applied to sparse arrays, and transformational array intrinsics applies to sparse arrays as well. ....

Tyng--Ruey Chuang, Rong--Guey Chang, and Jenq Kuen Lee. Sampling and analytical techniques for data distribution of parallel sparse computation. In Eighth SIAM Conference on Parallel Processing for Scientific Computing. Minneapolis, Minnestota, USA, March 1997. 8 pages. SIAM Press.

....on other domains of interests (i.e. non zero ratios, compression distribution schemes, cost functions, etc. is known as abstract interpretation [1] Several recent works have used statistical information about the applications or the target machines to help select implementation strategies [3] [6] [7] 8] 15] These methods seems to concentrate on specific applications, while ours centers around high level matrix classes for and dynamic selections of specialized classes for faster execution. We are currently applying our method to larger applications to show that the method is ....

Tyng--Ruey Chuang, Rong--Guey Chang, and Jenq Kuen Lee. Sampling and analytical techniques for data distribution of parallel sparse computation. In Eighth SIAM Conference on Parallel Processing for Scientific Computing. Minneapolis, Minnestota, March 1997. 8 pages. SIAM Press.

....execution. The selection depends on the non zero ratios of the matrices, on the structures of the matrix computation, and on the performance characteristics of the underlining execution environment. Similar analysis has been carried out in the context of Fortran 90 array intrinsic operators [4]. However, we emphasize that Java and its environment allow us to perform such analysis continuously at run4 time. This is contrary to our previous work on analyzing Fortran 90 array intrinsics for sparse computation, which is a compile time effort. For each supported matrix operation, and given ....

....non zero ratios, compression distribution schemes, cost functions, etc. is known as abstract interpretation in the functional programming literature. Several recent works have used statistical information about the applications or the target machines to help select implementation strategies [4, 5, 11]. These methods seem to concentrate on specific applications, while ours center around high level matrix classes and dynamic selections of specialized classes for faster execution. Runtime compilation techniques and their applications to sparse computation can be found at [12, 16] For the ....

Tyng--Ruey Chuang, Rong--Guey Chang, and Jenq Kuen Lee. Sampling and analytical techniques for data distribution of parallel sparse computation. In Eighth SIAM Conference on Parallel Processing for Scientific Computing. Minneapolis, Minnesota, USA, March 1997. 8 pages. SIAM Press.

....= norm(b,size(b) if (normr .le. tol normb) exit p = r beta p Ap = MATMUL(A,p) alpha = rtr DOTPRODUCT(p,Ap) x = x alpha p r = r alpha Ap rtrold = rtr rtr = DOTPRODUCT(r,r) beta = rtr rtrold end do 2. 2 Higher Dimension Compression and Distribution Schemes In recent research effort, such as [23, 7, 4], the distribution and compressed schemes being considered for sparse matrices are limited to one dimension and two dimension arrays. We consider here schemes for higher dimensional arrays as well. The distribution schemes currently being considered are general block partitions based on number of ....

....cost functions of those routines (one for each choice of representation) can be evaluated according to the current sparsity of the input matrices. The routine with the smallest cost will be selected as the representation. Detailed description of this approach can be found in our earlier work[7]. 5.4 Storage Reuse and Aggregate Data Access Look at the following Fortran 90 array statement, which is a part of the relaxation example as described in Section 4. u(2:n 1:2,2:n 1:2) 0.25 (u(3:n:2,2:n 1:2) u(1:n 2:2,2:n 1:2) u(2:n 1:2,3:n:2) u(2:n 1:2,1:n 2:2) h2 rhs(2:n 1:2,2:n 1:2) A ....

[Article contains additional citation context not shown here]

Tyng--Ruey Chuang, Rong--Guey Chang, and Jenq Kuen Lee. Sampling and analytical techniques for data distribution of parallel sparse computation. In Eighth SIAM Conference on Parallel Processing for Scientific Computing. Minneapolis, Minnesota, USA, March 1997. 8 pages. SIAM Press.

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