国家自然科学基金资助项目（60921062，61003087）

Abstract

Abstract:

The applicationlevel checkpointing technique is one of the most commonly used and well matured faulttolerance techniques in homogenous systems. However, It is on its infant phase in heterogeneous systems and there are not accurate and reasonable solutions or approaches with respect to architectures and fault models of heterogeneous systems. Motivated by this observation, based on the architecture and programming model of the CUDA heterogeneous system, this paper analyzes the execution mode of CUDA programs running on CPUs and GPUs and proposes an asynchronous execution mechanism for the applicationlevel checkpointing technique in heterogeneous systems. With this mechanism, we explore a solution of optimal placement of checkpoints in heterogeneous systems. Finally, three experimental cases in the CUDA platform are used to evaluate our technique’s performance, feasibility and viability. The results demonstrate the effectiveness of our asynchronous execution mechanism for the applicationlevel checkpointing technique in heterogeneous systems. Compared with the synchronous execution mechanism, our mechanism is more flexible and has broader optimization space to explore. Moreover, Our solution of optimal placement of checkpoints can efficiently reduce checkpointing overhead and hence obtain higher performance.

Key words: applicationlevel checkpointing;heterogeneous system;asynchronous execution mechanism;optimal placement of checkpoints

JIA Jia. 国家自然科学基金资助项目（60921062，61003087）[J]. J4, 2011, 33(11): 54-59.

References

［1］Luebke D, Harris M, Krüger J, et al.GPGPU: General Purpose Computation on Graphics Hardware［C］∥Proc of ACM SIGGRAPH’04,2004:33.
［2］Fan Z, Qiu F, Kaufman A, et al. GPU Cluster for High Performance Computing［C］∥Proc of the 2004 ACM/IEEE Conference on Supercomputing,2004:47.
［3］http://www.top500.org.
［4］Ramkumar B,Strumpen V. Portable Checkpointing for Heterogeneous Architectures［C］∥Proc of the 27th International Symposium on FaultTolerant Computing,1997:5867.
［5］Kirk D. NVIDIA CUDA Software and GPU Parallel Computing Architecture［C］∥Proc of the 6th International Symposium on Memory Management,2007:103104.
［6］Kapasi U J, Rixner S, Dally W J, et al. Programmable Stream Processors［J］. IEEE Computer, 2003,36(8):5462.
［7］Advanced Micro Devices, Inc. AMD Brook+［EB/OL］.［20091116］.http://ati.amd.com/technology/streamcomputing/AMDBrookplus.pdf.
［8］Open Computing Language［EB/OL］.［20091116］.http://www.khronos.org/.
［9］CUDA Technical Training Volume I/II［R］. Prepared and Provided by NVIDIA, 2008.
［10］Elnozahy E N, Alvisi L, Wang Y, et al. A Survey of RollbackRecovery Protocols in MessagePassing Systems［J］. ACM Computing Surveys, 2002,34(3):375408.
［11］Plank J S, Li K, Puening M A. Diskless Checkpointing［J］. IEEE Transactions Parallel and Distributed Systems, 1998,9(10):972986.
［12］Compute Visual Profiler 4.0 for NVIDIA CUDA User Guide［R］. DU05162001_v04, 2011.
［13］Sheaffer J, Luebke D, Skadron K. A Hardware Redundancy and Recovery Mechanism for Reliable Scientific Computation on Graphics Processors［C］∥Proc of Graphics Hardware, 2007:5564.
［14］Dimitrov M, Mantor M, Zhou H. Understanding Software Approaches for GPGPU Reliability［C］∥Proc of the 2nd Workshop on General Purpose Processing on Graphics Processing Units,2009:94104.
［15］Krishna C M, YannHang L, Kang G S. Optimization Criteria for Checkpoint Placement［J］. Communication of the ACM, 1984, 27(10):10081012.
［16］Chandy K M,Ramamoorthy C V. Rollback and Recovery Strategies for Computer Programs［J］. IEEE Transactions on Computers, 1972,21(6):546556.
［17］Toueg S, Babaolu . On the Optimum Checkpoint Selection Problem［J］. SIAM Journal on Computing, 1984,13(3):630649.
［18］Upadhyaya S J,Saluja K K. An Experimental Study to Determine Task Size for Rollback Recovery Systems［J］. IEEE Transactions on Computers, 1988,37(7):872877.