C. Ding and K. Kennedy. Inter-array data regrouping. In Proceedings of The 12th International Workshop on Languages and Compilers for Parallel Computing, La Jolla, California, August 1999.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....of nodes and edges in a representative graph are as follows. Data set nodes edges mol1 131072 1179648 mol2 442368 3981312 foil 144649 1074393 auto 448695 3314611 The baseline benchmarks and the executors for the runtime reordering transformation compositions use inter array data regrouping [8] to leverage shared memory reference patterns between data arrays. All compositions we consider consist of a data reordering transformation (CPACK or Gpart) followed by the iteration reordering transformation lexicographical grouping (lexGroup) for the j loop. We also perform the composition ....

C. Ding and K. Kennedy. Inter-array data regrouping. In Twelfth International Workshop on Languages and Compilers for Parallel Computing. Springer-Verlag, August 1999.

.... irregular computations and improve locality [1] Ding and Kennedy explored applying dynamic copying (packing) of data elements based on loop traversal order [22] They also developed algorithms for reorganizing single arrays into multi dimensional arrays depending on their access patterns [23]. Mellor Crummey et al. use space filling curves to map multidimensional data to memory [58] However, space filling curves must be applied by hand using geometric coordinate information. Mitchell et al. improved locality using bucket sorting to reorder loop iterations in irregular computations ....

C. Ding and K. Kennedy. Inter-array data regrouping. In Proceedings of the Twelfth Workshop on Languages and Compilers for Parallel Computing, San Diego, CA, August 1999.

....we show partitioning algorithms can yield better locality, albeit with higher processing overhead. Ding and Kennedy also developed algorithms for reorganizing single arrays into multi dimensional arrays depending on their access patterns, improving the performance of many irregular applications [9]. Their technique might be useful for Fortran codes where data are often single dimension arrays, not structures as in C codes. Mellor Crummey et al. used a geometric partitioning algorithm based on space filling curves to map multidimensional data to memory [27] They also blocked computation ....

C. Ding and K. Kennedy. Inter-array data regrouping. In Proceedings of the Twelfth Workshop on Languages and Compilers for Parallel Computing, San Diego, August 1999.

....sophisticated on the fly algorithm for deciding when to reapply transformations for adaptive codes. Ding and Kennedy also developed algorithms for reorganizing single arrays into multi dimensional arrays depending on their access patterns, improving the performance of many irregular applications [12]. Their technique might be useful for Fortran codes where data are often single dimension arrays, not structures as in C codes. Mellor Crummey et al. used a geometric partitioning algorithm based on space filling curves to map multidimensional data to memory [33] They also blocked computation ....

C. Ding and K. Kennedy. Inter-array data regrouping. In Proceedings of the Twelfth Workshop on Languages and Compilers for Parallel Computing, San Diego, August 1999.

....sophisticated on the fly algorithm for deciding when to reapply transformations for adaptive codes. Ding and Kennedy also developed algorithms for reorganizing single arrays into multi dimensional arrays depending on their access patterns, improving the performance of many irregular applications [12]. Their technique might be useful for Fortran codes where data are often single dimension arrays, not structures as in C codes. Mellor Crummey et al. used a geometric partitioning algorithm based on space filling curves to map multidimensional data to memory [34] They also blocked computation ....

C. Ding and K. Kennedy. Inter-array data regrouping. In Proceedings of the Twelfth Workshop on Languages and Compilers for Parallel Computing, San Diego, CA, August 1999.

....we show partitioning algorithms can yield better locality, albeit with higher processing overhead. Ding andKennedy also developed algorithms for reorganizing single arrays into multi dimensional arrays depending on their access patterns, improving the performance of many irregular applications [9]. Their technique might be useful for Fortran codes where data are often single dimension arrays, not structures as in C codes. Mellor Crummey et al. used a geometric partitioning algorithm based on space filling curves to map multidimensional data to memory [25] They also blocked computation ....

C. Ding and K. Kennedy. Inter-array data regrouping. In Proceedings of the Twelfth Workshop on Languages and Compilers for Parallel Computing, San Diego, August 1999.

.... to keep values associated with each node (e.g. weight, velocity, force, etc) One multi dimensional array may be used instead of several arrays, but only arrays frequently used together should be merged into a multi dimensional array to fully exploit memory bandwidth and cache line utilization [15]. For edge structures, two type of structures are commonly used; edge lists and partner lists. Figure 9 demonstrates how a graph is implemented using edge and partner lists, along with changes due to data and computation reordering. In the example, we store upper triangular part of edge list ....

....us to develop high quality but low overhead algorithms. Ding and Kennedy also discussed reorganization of single arrays into multi dimensional arrays depending on how closely they are accessed in computation, and found that their technique improves the performance of irregular applications [15]. They also showed their technique does not hurt the performance of regular applications if back end compilers are smart enough to extract the existing instruction level parallelism in transformed codes. Mellor Crummey et al. used a geometric ordering algorithm based on space filling curves to map ....

C. Ding and K. Kennedy. Inter-array data regrouping. In Proceedings of the Twelveth Workshop on Languages and Compilers for Parallel Computing, San Diego, Aug. 1999.

No context found.

C. Ding and K. Kennedy. Inter-array data regrouping. In Proceedings of The 12th International Workshop on Languages and Compilers for Parallel Computing, La Jolla, California, August 1999.

....time 1.66 sec 1.30 sec 1.01 sec Origin2000 memory reads 512 MB 256 MB 256 MB memory writebacks 256 MB 256 MB 128 MB Exemplar execution time 0.68 sec 0.25 sec 0. 19 sec Table 1: Effect of data regrouping on writeback reduction The general formulation and solution of data regrouping is given in[8]. As illustrated by the paper, the regrouping algorithm first divides a program into computation phases and partitions its arrays into compatible sets. Regrouping is applied within each compatible set by grouping only those arrays that are always accessed together. The above example uses the same ....

C. Ding and K. Kennedy. Inter-array data regrouping. In Proceedings of The 12th International Workshop on Languages and Compilers for Parallel Computing, August 1999.

....serious memory bandwidth bottleneck[1] At Rice University, we are developing a compiler strategy for ameliorating the effect of memory bandwidth bottleneck. The first part is to minimize the overall amount of memory transfer through automatic compiler enhancement of global and dynamic cache reuse[1, 2, 3]. Although effective, automatic optimizations are not perfect both because they may fail in some cases and because they do not estimate program execution time, which is important for subsequent task scheduling. To overcome these problems, this paper presents the second part of this compiler ....

....the tool and the compiler. First, the tool should be aware of certain compiler transformations such as data reuse optimizations because they may change the actual amount of memory transfer. The most notable is the global transformations such as reuse based loop fusion[1] and global data regrouping[2], which can radically change the structure of both the computation and data of a program and can reduce the overall amount of memory transfer by integral factors. The performance tool must know these high level transformations for it to obtain an accurate estimate of memory transfer. In addition ....

[Article contains additional citation context not shown here]

C. Ding and K. Kennedy. Inter-array data regrouping. In Proceedings of The 12th International Workshop on Languages and Compilers for Parallel Computing, August 1999.

....two problems. The first is to analyze data structures of different sizes and different access patterns. The second, more difficult one, is to avoid undesirable side effects of a data transformation. The program analysis for data regrouping has been reported in an earlier workshop paper of ours[5], so has the algorithm for array regrouping on a single data dimension. Data regrouping first partitions a program into a sequence of computation phases, each of which accesses data that is larger than cache. Then, it classifies all data arrays into compatible groups. Two arrays are compatible if ....

....In a recent study, Ding and Kennedy used array shrinking and peeling to perform write back reduction after loop fusion and before data regrouping[6] Neither work has been tested on non trivial programs. Data regrouping is related to many data transformations, which have been discussed in [5]. Among those, data regrouping is the first one to selectively combine multiple arrays with guaranteed profitability and static optimality. And in this work, data regrouping is extended to grouping at multiple levels of high dimensional data. 6 Contributions This work has developed a global ....

C. Ding and K. Kennedy. Inter-array data regrouping. In Proceedings of The 12th International Workshop on Languages and Compilers for Parallel Computing, August 1999.

No context found.

C. Ding and K. Kennedy. Inter-array data regrouping. In Proceedings of The 12th International Workshop on Languages and Compilers for Parallel Computing, August 1999.

No context found.

C. Ding and K. Kennedy. Inter-array data regrouping. In Proc. the Workshop on Languages and Compilers for Parallel Computing, San Diego, CA, August 4--6, 1999.

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