J. Darlington, Y. Guo, and H.W. To. Structured Parallel Programming: Theory meets Practice. In Research Directions in Computer Science. Cambridge University Press, 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....compared to the row clustered version is shown by the smaller number of global pointers in the former: 137 vs. 387 on average for 16 processors. 5 Related Work Our clustering approach shows strong similarities to skeleton based parallel programming [16] exempli ed by systems such as SCL [17], P3L [18] SkelML [19] or Skil [20] with an emphasis on compositionality and genericity. Compared to skeletons we trade an optimized parallel implementation of some higher order functions for a more exible framework where new coordination constructs, and their granularity and locality can be ....

J. Darlington, Y. Guo, and H.W. To. Structured Parallel Programming: Theory meets Practice. In Research Directions in Computer Science. Cambridge University Press, 1996.

....The compiler applies source level transformations to optimize the parallel behaviour. This includes granularity adjustment, nested parallelism flattening, optimization of data distribution movement and interprocessor communications. This is all done using algebraic transformations as specified in [11]. To [23] discusses a skeleton based system allowing composition. The system provides a set of data parallel skeletons modeling collective operations over two parallel data types rLists (modeling restricted lists without a construct operator) and Arrays. Data parallel skeletons can be glued ....

J. Darlington, Y.K. Guo, and H. W. To. Structured parallel programming: Theory meets practice. Draft, unpublished, January 1995.

....augmented with message passing libraries. Due to the functional nature of SCL source level transformation can be applied to optimise the parallel behaviour of the program, including granularity adjustment, nested parallelism flattening, optimised data distribution and interprocessor communication [2]. Currently, we are building a prototype system based on Fortran 77 plus MPI [7] targeted at a Fujitsu AP1000 machine [6] The matrix multiplication example has been translated to Fortran77 plus MPI on an AP1000. Due to the richness of information provided by the Fortran S code, the performance ....

J. Darlington, Y. Guo, and H. W. To. Structured parallel programming: Theory meets practice. Technical report, Imperial College, 1995. unpublished.

....For example the rules: send f ffi send g = send (f ffi g) fetch f ffi fetch g = fetch (g ffi f) enable communication steps to be removed by combining two communication steps into one. By taking these transformation rules as algebraic laws, a powerful communication algebra is developed [6]. Flattening: Transformation can be applied to flatten nested data parallelism. Let sgf = gf 0 ffi map gf 1 ffi (split P) then the rule for flattening is: SPMD [ gf 0 , SPMD [ gf 1 , lf 1 ) ffi (split P) SPMD [ sgf, lf) With this rule, nested SPMD computation can be transformed into a ....

....rule, nested SPMD computation can be transformed into a flat data parallel computation with a segmented global function sgf. Thus, the sgf provides the a similar functionality to the Segmented Instructions [1] used in the NESL language implementation. More transformation rules can be found in [6]. 5 Experimental Results To assess the performance of code resulting from this approach some initial experiments have been performed. The experiments were conducted on a Fujitsu AP1000 [12] using Fortran and MPI as the target code for our hand compilation. MPI was chosen to ensure future ....

J. Darlington, Y. Guo, and H. W. To. Structured parallel programming: Theory meets practice. Technical report, Imperial College, 1995. unpublished.

....send f ffi send g = send (f ffi g) fetch f ffi fetch g = fetch (g ffi f) enable us to remove a communication step by combining two communication steps into one. By taking these transformation rules for optimising data communication as algebraic laws, a powerful communication algebra is developed [4]. Flattening: Transformation can be applied to flatten nested data parallelism. Let sgf = gf 0 ffi map gf 1 ffi (split P) then the rule for flattening is: SPMD [ gf 0 , SPMD [ gf 1 , lf 1 ) ffi (split P) SPMD [ sgf, lf) With the rule, nested SPMD computation can be transformed into a ....

....the rule, nested SPMD computation can be transformed into a flat data parallel computation with a segmented global function sgf. Thus, the sgf provides a similar functionality to the Segmented Instructions [13] used in the NESL language implementation. More transformation rules can be found in [4]. 5 Conclusion In this paper we have proposed functional skeletons as a new mechanism of developing general purpose parallel coordination systems. The work stems from our original work on functional skeletons to capture re occurring patterns of parallel computation. This has been extended such ....

J. Darlington, Y. Guo, H. W. To. Structured Parallel Programming: Theory meets Practice. To be appeared in the New Direction of Computing.

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