Reinforcement learning for automated performance tuning: Initial evaluation for sparse matrix format selection

Armstrong, Warren; Rendell, Alistair P.

doi:10.1109/clustr.2008.4663802

Cited by 5 publications

(6 citation statements)

References 9 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In this respect further investigation of the sensitivity of the BPS to the characteristics of the test matrices and the capabilities of the underlying hardware may be advantageous, as might be the use of machine learning techniques to provide a more sophisticated mapping function. It may also be possible to use reinforcement learning techniques that do not rely on pre-learning a set of rules to develop a BPS that can dynamically adapt to changes in runtime conditions [15].…”

Section: Discussionmentioning

confidence: 99%

Generating optimal CUDA sparse matrix–vector product implementations for evolving GPU hardware

Zein

Rendell

2011

Concurrency and Computation

Self Cite

View full text Add to dashboard Cite

SUMMARYThe CUDA model for graphics processing units (GPUs) presents the programmer with a plethora of different programming options. These includes different memory types, different memory access methods and different data types. Identifying which options to use and when is a non-trivial exercise. This paper explores the effect of these different options on the performance of a routine that evaluates sparse matrixvector products (SpMV) across three different generations of NVIDIA GPU hardware. A process for analysing performance and selecting the subset of implementations that perform best is proposed. The potential for mapping sparse matrix attributes to optimal CUDA SpMV implementations is discussed.

show abstract

Section: Discussionmentioning

confidence: 99%

Generating optimal CUDA sparse matrix–vector product implementations for evolving GPU hardware

Zein

Rendell

2011

Concurrency and Computation

Self Cite

View full text Add to dashboard Cite

show abstract

“…Reinforcement learning was proposed for sparse matrix format selection by Armstrong and Rendell [8]. This current work differs in the learning algorithm employed, the number of formats available (65 rather than 3) and the matrices used for evaluation.…”

Section: Related Workmentioning

confidence: 98%

Runtime sparse matrix format selection

Armstrong

Rendell

2010

Procedia Computer Science

Self Cite

View full text Add to dashboard Cite

There exist many storage formats for the in-memory representation of sparse matrices. Choosing the format that yields the quickest processing of any given sparse matrix requires considering the exact non-zero structure of the matrix, as well as the current execution environment. Each of these factors can change at runtime. The matrix structure can vary as computation progresses, while the environment can change due to varying system load, the live migration of jobs across a heterogeneous cluster, etc. This paper describes an algorithm that learns at runtime how to map sparse matrices onto the format which provides the quickest sparse matrix-vector product calculation, and which can adapt to the hardware platform changing underfoot. We show multiplication times reduced by over 10% compared with the best non-adaptive format selection.

show abstract

“…For matrices that have affinity to DIA, ELL or COO format (No. [1][2][3][4][5][6][7][8], the corresponding SpMVs achieve higher performance than those in CSR format (No.9-12). The performance gap indicates that it is meaningful to implement a high performance SpMV library being aware of sparse structures (applications).…”

Section: Performancementioning

confidence: 99%

“…Second, we extract more features from real matrices in UF collection, which can feed more training data to data mining tools so as to generate more reliable rules for the learning model. W. Armstrong et al [5] uses reinforcement learning to choose the best format, but users should decide the factor values which influence the accuracy of the learning model. SMAT system is more convenient to automatically generate the model and still achieve similar prediction accuracy.…”

Section: Related Workmentioning

confidence: 99%

“…All these libraries including Intel MKL [1] and OSKI [31] provide different interfaces for every format supported, which leaves the burden of determining and integrating the optimal format to application programmers. As a consequence, although these libraries are highly optimized by either hand-tuning or autotuning techniques [4,5,7,8,10,11,17,21,29,31,35,36] with processor architectures evolving, it is still difficult to integrate them into high level solvers or applications. This fact results in an embarrassing situation: on one hand, there are several high performance sparse libraries available; on the other hand, solver or application developers seldom use them in practice.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Smat

Tan

Chen

et al. 2013

Proceedings of the 34th ACM SIGPLAN Conference on Programming Language Design and Implementation

View full text Add to dashboard Cite

Sparse Matrix Vector multiplication (SpMV) is an important kernel in both traditional high performance computing and emerging data-intensive applications. By far, SpMV libraries are optimized by either application-specific or architecture-specific approaches, making the libraries become too complicated to be used extensively in real applications. In this work we develop a Sparse Matrixvector multiplication Auto-Tuning system (SMAT) to bridge the gap between specific optimizations and general-purpose usage. S-MAT provides users with a unified programming interface in compressed sparse row (CSR) format and automatically determines the optimal format and implementation for any input sparse matrix at runtime. For this purpose, SMAT leverages a learning model, which is generated in an off-line stage by a machine learning method with a training set of more than 2000 matrices from the UF sparse matrix collection, to quickly predict the best combination of the matrix feature parameters. Our experiments show that SMAT achieves impressive performance of up to 51GFLOPS in single-precision and 37GFLOPS in double-precision on mainstream x86 multi-core processors, which are both more than 3 times faster than the Intel MKL library. We also demonstrate its adaptability in an algebraic multigrid solver from Hypre library with above 20% performance improvement reported.

show abstract

Reinforcement learning for automated performance tuning: Initial evaluation for sparse matrix format selection

Cited by 5 publications

References 9 publications

Generating optimal CUDA sparse matrix–vector product implementations for evolving GPU hardware

Generating optimal CUDA sparse matrix–vector product implementations for evolving GPU hardware

Runtime sparse matrix format selection

Smat

Contact Info

Product

Resources

About