In the field of HPC, the current hardware trend is to design multiprocessor architectures featuring heterogeneous technologies such as specialized coprocessors (e.g. Cell/BE) or data-parallel accelerators (e.g. GPUs). Approaching the theoretical performance of these architectures is a complex issue. Indeed, substantial efforts have already been devoted to efficiently offload parts of the computations. However, designing an execution model that unifies all computing units and associated embedded memory remains a main challenge. We therefore designed StarPU, an original runtime system providing a high-level, unified execution model tightly coupled with an expressive data management library. The main goal of StarPU is to provide numerical kernel designers with a convenient way to generate parallel tasks over heterogeneous hardware on the one hand, and easily develop and tune powerful scheduling algorithms on the other hand. We have developed several strategies that can be selected seamlessly at run-time, and we have analyzed their efficiency on several algorithms running simultaneously over multiple cores and a GPU. In addition to substantial improvements regarding execution times, we have obtained consistent superlinear parallelism by actually exploiting the heterogeneous nature of the machine. We eventually show that our dynamic approach competes with the highly optimized MAGMA library and overcomes the limitations of the corresponding static scheduling in a portable way. Copyright © 2010 John Wiley & Sons, Ltd.

/pdf/starpu-a-unified-platform-for-task-scheduling-on-2qdx6n5x2r.pdf

StarPU: a unified platform for task scheduling on heterogeneous multicore architectures

Sparse matrix-vector multiplication (SpMV) is of singular importance in sparse linear algebra. In contrast to the uniform regularity of dense linear algebra, sparse operations encounter a broad spectrum of matrices ranging from the regular to the highly irregular. Harnessing the tremendous potential of throughput-oriented processors for sparse operations requires that we expose substantial fine-grained parallelism and impose sufficient regularity on execution paths and memory access patterns. We explore SpMV methods that are well-suited to throughput-oriented architectures like the GPU and which exploit several common sparsity classes. The techniques we propose are efficient, successfully utilizing large percentages of peak bandwidth. Furthermore, they deliver excellent total throughput, averaging 16 GFLOP/s and 10 GFLOP/s in double precision for structured grid and unstructured mesh matrices, respectively, on a GeForce GTX 285. This is roughly 2.8 times the throughput previously achieved on Cell BE and more than 10 times that of a quad-core Intel Clovertown system.

/pdf/implementing-sparse-matrix-vector-multiplication-on-2m98krk6l7.pdf

Implementing sparse matrix-vector multiplication on throughput-oriented processors

The massive parallelism of graphics processing units (GPUs) oers tremendous performance in many high-performance computing applications. While dense linear algebra readily maps to such platforms, harnessing this potential for sparse matrix computations presents additional challenges. Given its role in iterative methods for solving sparse linear systems and eigenvalue problems, sparse matrix-vector multiplication (SpMV) is of singular importance in sparse linear algebra. In this paper we discuss data structures and algorithms for SpMV that are eciently implemented on the CUDA platform for the ne-grained parallel architecture of the GPU. Given the memory-bound nature of SpMV, we emphasize memory bandwidth eciency and compact storage formats. We consider a broad spectrum of sparse matrices, from those that are well-structured and regular to highly irregular matrices with large imbalances in the distribution of nonzeros per matrix row. We develop methods to exploit several common forms of matrix structure while oering alternatives which accommodate greater irregularity. On structured, grid-based matrices we achieve performance of 36 GFLOP/s in single precision and 16 GFLOP/s in double precision on a GeForce GTX 280 GPU. For unstructured nite-element matrices, we observe performance in excess of 15 GFLOP/s and 10 GFLOP/s in single and double precision respectively. These results compare favorably to prior state-of-the-art studies of SpMV methods on conventional multicore processors. Our double precision SpMV performance is generally two and a half times that of a Cell BE with 8 SPEs and more than ten times greater than that of a quad-core Intel Clovertown system.

/pdf/ecient-sparse-matrix-vector-multiplication-on-cuda-421t1ah7xg.pdf

Ecient Sparse Matrix-Vector Multiplication on CUDA

We present performance results for dense linear algebra using recent NVIDIA GPUs. Our matrix-matrix multiply routine (GEMM) runs up to 60% faster than the vendor's implementation and approaches the peak of hardware capabilities. Our LU, QR and Cholesky factorizations achieve up to 80--90% of the peak GEMM rate. Our parallel LU running on two GPUs achieves up to ~540 Gflop/s. These results are accomplished by challenging the accepted view of the GPU architecture and programming guidelines. We argue that modern GPUs should be viewed as multithreaded multicore vector units. We exploit blocking similarly to vector computers and heterogeneity of the system by computing both on GPU and CPU. This study includes detailed benchmarking of the GPU memory system that reveals sizes and latencies of caches and TLB. We present a couple of algorithmic optimizations aimed at increasing parallelism and regularity in the problem that provide us with slightly higher performance.

/pdf/benchmarking-gpus-to-tune-dense-linear-algebra-5rvwh9lyb0.pdf

Benchmarking GPUs to tune dense linear algebra

/pdf/benchmarking-gpus-to-tune-dense-linear-algebra-217tluukfj.pdf

Current machine translation (MT) systems are still not perfect. In practice, the output from these systems needs to be edited to correct errors. A way of increasing the productivity of the whole translation process (MT plus human work) is to incorporate the human correction activities within the translation process itself, thereby shifting the MT paradigm to that of computer-assisted translation. This model entails an iterative process in which the human translator activity is included in the loop: In each iteration, a prefix of the translation is validated (accepted or amended) by the human and the system computes its best (or n-best) translation suffix hypothesis to complete this prefix. A successful framework for MT is the so-called statistical (or pattern recognition) framework. Interestingly, within this framework, the adaptation of MT systems to the interactive scenario affects mainly the search process, allowing a great reuse of successful techniques and models. In this article, alignment templates, phrase-based models, and stochastic finite-state transducers are used to develop computer-assisted translation systems. These systems were assessed in a European project (TransType2) in two real tasks: The translation of printer manuals; manuals and the translation of the Bulletin of the European Union. In each task, the following three pairs of languages were involved (in both translation directions): English-Spanish, English-German, and English-French.

/pdf/statistical-approaches-to-computer-assisted-translation-4n8ioadd27.pdf

Statistical approaches to computer-assisted translation

The increase in performance of the last generations of graphics processors (GPUs) has made this class of platform a coprocessing tool with remarkable success in certain types of operations. In this paper we evaluate the performance of the Level 3 operations in CUBLAS, the implementation of BIAS for NVIDIAreg GPUs with unified architecture. From this study, we gain insights on the quality of the kernels in the library and we propose several alternative implementations that are competitive with those in CUBLAS. Experimental results on a GeForce 8800 Ultra compare the performance of CUBLAS and the new variants.

/pdf/evaluation-and-tuning-of-the-level-3-cublas-for-graphics-3q7jxl2tj2.pdf

Evaluation and tuning of the Level 3 CUBLAS for graphics processors

Some approaches to statistical and finite-state speech-to-speech translation

We present several algorithms to compute the solution of a linear system of equations on a GPU, as well as general techniques to improve their performance, such as padding and hybrid GPU-CPU computation. We also show how iterative refinement with mixed-precision can be used to regain full accuracy in the solution of linear systems. Experimental results on a G80 using CUBLAS 1.0, the implementation of BLAS for NVIDIA® GPUs with unified architecture, illustrate the performance of the different algorithms and techniques proposed.

/pdf/solving-dense-linear-systems-on-graphics-processors-1mt7ju4zyx.pdf

Solving Dense Linear Systems on Graphics Processors

We present several algorithms to compute the solution of a linear system of equations on a graphics processor (GPU), as well as general techniques to improve their performance, such as padding and hybrid GPU-CPU computation. We compare single and double precision performance of a modern GPU with unified architecture, and show how iterative refinement with mixed precision can be used to regain full accuracy in the solution of linear systems, exploiting the potential of the processor for single precision arithmetic. Experimental results on a GTX280 using CUBLAS 2.0, the implementation of BLAS for NVIDIA® GPUs with unified architecture, illustrate the performance of the different algorithms and techniques proposed. Copyright © 2009 John Wiley & Sons, Ltd.

/pdf/exploiting-the-capabilities-of-modern-gpus-for-dense-matrix-9kad5uj5ai.pdf

Sergio Barrachina

Papers

Statistical approaches to computer-assisted translation

Evaluation and tuning of the Level 3 CUBLAS for graphics processors

Some approaches to statistical and finite-state speech-to-speech translation

Solving Dense Linear Systems on Graphics Processors

Exploiting the capabilities of modern GPUs for dense matrix computations