Automatic Data Distribution for Improving Data Locality on the Cell BE Architecture

Wang, Miao; Bodin, François; Matz, Sébastien

doi:10.1007/978-3-642-13374-9_17

Cited by 3 publications

(3 citation statements)

References 13 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…• A single source compiler [9] processes each codelet and generates multiple variants. For each codelet, the compiler creates a pair of C source files -one for PPE and one for the SPEs -and the datasets' partitioning description.…”

Section: A Experimental Frameworkmentioning

confidence: 99%

“…We have designed a single source compiler [9] to generate SPMD codes for the codelet on the Cell BE. Our compiler first determines data alignment based on the Component Af f inity Graph (CAG).…”

Section: Data-parallel Code Generation For Codeletsmentioning

confidence: 99%

“…Each computational phase embeds tasks that will be executed concurrently, while the communication phase enables data transmission between successive computational phases. The data parallelism in each task is exploited using a single source compiler [9] under the SPMD (Single Program Multiple Data) model. In order to determine the optimal parallelization schemes for the multi-task programs, our model constructs a three-dimensional optimization space and explores it to find out the efficient schemes.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Model-Directed Multi-Level Parallelization for Multi-Task Programs on the Cell BE

Miao

Benoît²,

Bodin³

2009

2009 2nd International Conference on Computer Science and Its Applications

Self Cite

View full text Add to dashboard Cite

Multi-core architectures can deliver high processing power if the multiple levels of parallelism they expose are exploited. However, it is non-trivial to orchestrate the computational and memory resources allocation. Furthermore, when dealing with distributed memory architectures, data distribution adds another level of complexity. This paper presents a modeldriven technique for mapping multi-task parallel programs onto multi-core platforms. The resource allocation configurations are expressed within a three-dimensional optimization space which affects task-level and data-level parallelism, and communication. Then, the performance of any valid parallelization scheme is estimated statically, taking into account both computation and communication costs. We prototyped our approach on the Cell BE with an image processing application. Experiments show that our model can correctly predict the overall performance and highlight the most efficient parallelization schemes.

show abstract

Section: A Experimental Frameworkmentioning

confidence: 99%

Section: Data-parallel Code Generation For Codeletsmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation