Performance optimization methods for boosting MIC-based parallel computing
XU Zhiliang · Journal of Shenzhen Institute of Information Technology · 2013
To satisfy the ever-increasing performance demand of many integrated core(MIC) based parallel computing and critical application's operation,the performance optimization methods needs to offer efficient latency hiding,flexible data transfer,and high scalability features so as to facilitate and simplify the parallelism development.Benefits from the same application code base,MIC could amplify the parallelism from applications and accelerate them as tasks with SIMD kernels.The in-memory database(IMDb) eliminates the I/O bottleneck by storing data in main memory.We give a unified approach to fully utilize the advantage of MIC for IMDb online transaction,a threelevel optimization design strategy,including the memory-access level,kernel-speedup level and data-partition level,is also proposed using the hardware parallelism to achieve task-level and data-level parallelism of IMDb programs,and guarantee that the IMDb could express real-time transaction in an efficient way.