The slowing pace of commodity microprocessor performance improvements combined with ever-increasing chip power demands has become of utmost concern to computational scientists. As a result, the high performance computing community is examining alternative architectures that address the limitations of modern cache-based designs. In this work, we examine the potential of using the forthcoming STI Cell processor as a building block for future high-end computing systems. Our work contains several novel contributions. First, we introduce a performance model for Cell and apply it to several key scientific computing kernels: dense matrix multiply, sparse matrix vector multiply, stencil computations, and 1D/2D FFTs. The difficulty of programming Cell, which requires assembly level intrinsics for the best performance, makes this model useful as an initial step in algorithm design and evaluation. Next, we validate the accuracy of our model by comparing results against published hardware results, as well as our own implementations on the Cell full system simulator. Additionally, we compare Cell performance to benchmarks run on leading superscalar (AMD Opteron), VLIW (Intel Itanium2), and vector (Cray X1E) architectures. Our work also explores several different mappings of the kernels and demonstrates a simple and effective programming model for Cell’s unique architecture. Finally, we propose modest microarchitectural modifications that could significantly increase the efficiency of double-precision calculations. Overall results demonstrate the tremendous potential of the Cell architecture for scientific computations in terms of both raw performance and power efficiency.

Stencil-based kernels constitute the core of many scientific applications on block-structured grids. Unfortunately, these codes achieve a low fraction of peak performance, due primarily to the disparity between processor and main memory speeds. We examine several optimizations on both the conventional cache-based memory systems of the Itanium 2, Opteron, and Power5, as well as the heterogeneous multicore design of the Cell processor. The optimizations target cache reuse across stencil sweeps, including both an implicit cache oblivious approach and a cache-aware algorithm blocked to match the cache structure. Finally, we consider stencil computations on a machine with an explicitlymanaged memory hierarchy, the Cell processor. Overall, results show that a cache-aware approach is significantly faster than a cache oblivious approach and that the explicitly managed memory on Cell is more efficient: Relative to the Power5, it has almost 2x more memory bandwidth and is 3.7x faster. 1.

Stencil-based kernels constitute the core of many important scientific applications on block-structured grids. Unfortunately, these codes achieve a low fraction of peak performance, due primarily to the disparity between processor and main memory speeds. In this paper, we explore the impact of trends in memory subsystems on a variety of stencil optimization techniques and develop performance models to analytically guide our optimizations. Our work targets cache reuse methodologies across single and multiple stencil sweeps, examining cache-aware algorithms as well as cache-oblivious techniques on the Intel Itanium2, AMD Opteron, and IBM Power5. Additionally, we consider stencil computations on the heterogeneous multi-core design of the Cell processor, a machine with an explicitly-managed memory hierarchy. Overall our work represents one of the most extensive analyses of stencil optimizations and performance modeling to date. Results demonstrate that recent trends in memory system organization have reduced the efficacy of traditional cacheblocking optimizations. We also show that a cache-aware implementation is significantly faster than a cache-oblivious approach, while the explicitly managed memory on Cell enables the highest overall efficiency: Cell attains 88 % of algorithmic peak while the best competing cache-based processor only achieves 54 % of algorithmic peak performance.

The slowing pace of commodity microprocessor performance improvements combined with ever-increasing chip power demands has become of utmost concern to computational scientists. As a result, the high performance computing community is examining alternative architectures that address the limitations of modern cache-based designs. In this work, we examine the potential of using the recently-released STI Cell processor as a building block for future high-end computing systems. Our work contains several novel contributions. First, we introduce a performance model for Cell and apply it to several key scientific computing kernels: dense matrix multiply, sparse matrix vector multiply, stencil computations, and 1D/2D FFTs. The difficulty of programming Cell, which requires assembly level intrinsics for the best performance, makes this model useful as an initial step in algorithm design and evaluation. Next, we validate the accuracy of our model by comparing results against published hardware results, as well as our own implementations on a 3.2GHz Cell blade. Additionally, we compare Cell performance to benchmarks run on leading superscalar (AMD Opteron), VLIW (Intel Itanium2), and vector (Cray X1E) architectures. Our work also explores several different mappings of the kernels and demonstrates a simple and effective programming model for Cell’s unique architecture. Finally, we propose modest microarchitectural modifications that could significantly increase the efficiency of double-precision calculations. Overall results demonstrate the tremendous potential of the Cell architecture for scientific computations in terms of both raw performance and power efficiency. 1.

Stencil computations form the performance-critical core of many applications. Tiling and parallelization are two important optimizations to speed up stencil computations. Many tiling and parallelization strategies are applicable to a given stencil computation. The best strategy depends not only on the combination of the two techniques, but also on many parameters: tile and loop sizes in each dimension; computation-communication balance of the code; processor architecture; message startup costs; etc. The best choices can only be determined through design-space exploration, which is extremely tedious and error prone to do via exhaustive experimentation. We characterize the space of multi-level tilings and parallelizations for 2D/3D Gauss-Siedel stencil computation. A systematic exploration of a part of this space enabled us to derive a design which is up to a factor of two faster than the standard implementation. 1.

All rights reserved.

Abstract not found

We present a new cache oblivious scheme for iterative stencil computations that performs beyond system bandwidth limitations as though gigabytes of data could reside in an enormous on-chip cache. We compare execution times for 2D and 3D spatial domains with up to 128 million double precision elements for constant and variable stencils against hand-optimized naive code and the automatic polyhedral parallelizer and locality optimizer PluTo and demonstrate the clear superiority of our results. The performance benefits stem from a tiling structure that caters for data locality, parallelism and vectorization simultaneously. Rather than tiling the iteration space from inside, we take an exterior approach with a predefined hierarchy, simple regular parallelogram tiles and a locality preserving parallelization. These advantages come at the cost of an irregular work-load distribution but a tightly integrated loadbalancer ensures a high utilization of all resources.

Abstract not found

Microprocessor speed has been growing exponentially faster than memory speed in the recent past. In this paper, we consider the long term implications of a continuation of this trend. We define a program property known as scalable locality, which measures our ability to apply ever faster uniprocessors to increasingly large problems (just as scalable parallelism measures our ability to apply more numerous processors to larger problems). We then show that exploiting scalable locality in scientific programs often requires advanced compile-time reordering of loop iterations. We provide an algorithm that derives an execution order, storage mapping, and cache requirement for any desired degree of locality, for certain programs that can be made to exhibit scalable locality. Our approach is unusual in that it derives the program transformation and cache requirement from the dataow of the calculation (a fundamental characteristic of the algorithm), instead of searching a space of possible transformations of the execution order and array layout used by the programmer (which are artifacts of the expression of the algorithm). We include empirical results showing the effectiveness of our transformation on two small kernels and a non-trivial benchmark program. Our transformation can produce speedups for data sets residing in L2 cache, main memory, or virtual memory. This report unifies the major results of DCS-TR-379 and DCS-TR-378, presents them in terms of a general algorithm rather than a collection of heuristics, and gives a clearer presentation of the empirical data.