TiNy Threads on BlueGene/P: Exploring Many-Core Parallelisms Beyond The Traditional OS
Uloženo v:
| Název: | TiNy Threads on BlueGene/P: Exploring Many-Core Parallelisms Beyond The Traditional OS |
|---|---|
| Autoři: | Ong Ye, Robert Pavel, Aaron L, Guang R. Gao |
| Přispěvatelé: | The Pennsylvania State University CiteSeerX Archives |
| Zdroj: | http://www.capsl.udel.edu/pub/doc/papers/MTAAP-2010.pdf. |
| Sbírka: | CiteSeerX |
| Popis: | Operating Systems have been considered as a cor-nerstone of the modern computer system, and the con-ventional operating system model targets computers designed around the sequential execution model. How-ever, with the rapid progress of the multi-core/many-core technologies, we argue that OSes must be adapted to the underlying hardware platform to fully exploit parallelism. To illustrate this, our paper reports a study on how to perform such an adaptation for the IBM BlueGene/P multi-core system. This paper’s major contributions are threefold. First, we have proposed a strategy to isolate the traditional OS functions to a single core of the BG/P multi-core chip, leaving the management of the remaining cores to a runtime software that is optimized to realize the parallel semantics of the user application according to a parallel program execution model. Second, we have ported the TNT (TiNy Thread) execution model to allow for further utilization of the BG/P compute cores. Finally, we have expanded the design framework described above to a multi-chip system designed for scalability to a large number of chips. An implementation of our method has been com-pleted on the Surveyor BG/P machine operated by Ar-gonne National Laboratory. Our experimental results provide insight into the strengths of this approach: (1) The performance of the TNT thread system shows comparable speedup to that of Pthreads running on the same hardware; (2) The distributed shared memory operates at 95 % of the experimental peak performance of the machine, with distance between nodes not being a sensitive factor; (3) The cost of thread creation shows a linear relationship as threads increase; (4) The cost of waiting at a barrier is constant and independent of the number of threads involved. 1. |
| Druh dokumentu: | text |
| Popis souboru: | application/pdf |
| Jazyk: | English |
| Relation: | http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.653.2372 |
| Dostupnost: | http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.653.2372 http://www.capsl.udel.edu/pub/doc/papers/MTAAP-2010.pdf |
| Rights: | Metadata may be used without restrictions as long as the oai identifier remains attached to it. |
| Přístupové číslo: | edsbas.B4873D51 |
| Databáze: | BASE |
Nájsť tento článok vo Web of Science | Abstrakt: | Operating Systems have been considered as a cor-nerstone of the modern computer system, and the con-ventional operating system model targets computers designed around the sequential execution model. How-ever, with the rapid progress of the multi-core/many-core technologies, we argue that OSes must be adapted to the underlying hardware platform to fully exploit parallelism. To illustrate this, our paper reports a study on how to perform such an adaptation for the IBM BlueGene/P multi-core system. This paper’s major contributions are threefold. First, we have proposed a strategy to isolate the traditional OS functions to a single core of the BG/P multi-core chip, leaving the management of the remaining cores to a runtime software that is optimized to realize the parallel semantics of the user application according to a parallel program execution model. Second, we have ported the TNT (TiNy Thread) execution model to allow for further utilization of the BG/P compute cores. Finally, we have expanded the design framework described above to a multi-chip system designed for scalability to a large number of chips. An implementation of our method has been com-pleted on the Surveyor BG/P machine operated by Ar-gonne National Laboratory. Our experimental results provide insight into the strengths of this approach: (1) The performance of the TNT thread system shows comparable speedup to that of Pthreads running on the same hardware; (2) The distributed shared memory operates at 95 % of the experimental peak performance of the machine, with distance between nodes not being a sensitive factor; (3) The cost of thread creation shows a linear relationship as threads increase; (4) The cost of waiting at a barrier is constant and independent of the number of threads involved. 1. |
|---|