Programming the CELL-BE using CSP

TY - CHAP

T1 - Programming the CELL-BE using CSP

AU - Skovhede, Kenneth

AU - Larsen, Morten N

AU - Vinter, Brian

PY - 2011

Y1 - 2011

N2 - The current trend in processor design seems to focus on using multiple cores, similar to a cluster-on-a-chip model. These processors are generally fast and power efficient, but due to their highly parallel nature, they are notoriously difficult to program for most scientists. One such processor is the CELL broadband engine (CELL-BE) which is known for its high performance, but also for a complex programming model which makes it difficult to exploit the architecture to its full potential. To address this difficulty, this paper proposes to change the programming model to use the principles of CSP design, thus making it simpler to program the CELL-BE and avoid livelocks, deadlocks and race conditions. The CSP model described here comprises a thread library for the synergistic processing elements (SPEs) and a simple channel based communication interface. To examine the scalability of the implementation, experiments are performed with both scientific computational cores and synthetic workloads. The implemented CSP model has a simple API and is shown to scale well for problems with significant computational requirements.

AB - The current trend in processor design seems to focus on using multiple cores, similar to a cluster-on-a-chip model. These processors are generally fast and power efficient, but due to their highly parallel nature, they are notoriously difficult to program for most scientists. One such processor is the CELL broadband engine (CELL-BE) which is known for its high performance, but also for a complex programming model which makes it difficult to exploit the architecture to its full potential. To address this difficulty, this paper proposes to change the programming model to use the principles of CSP design, thus making it simpler to program the CELL-BE and avoid livelocks, deadlocks and race conditions. The CSP model described here comprises a thread library for the synergistic processing elements (SPEs) and a simple channel based communication interface. To examine the scalability of the implementation, experiments are performed with both scientific computational cores and synthetic workloads. The implemented CSP model has a simple API and is shown to scale well for problems with significant computational requirements.

M3 - Book chapter

SP - 55

EP - 70

BT - Certified Associate Programmer

ER -