MIC商用并行编程性能优化分析

Performance optimization methods for boosting MIC-based parallel computing

针对MIC架构处理各种复杂业务时对性能日益增长的要求,为充分利用MIC使用已有编程模型的优势,通过避免内存容量、网络带宽方面的瓶颈增强并行编程的线程扩展性,对并行度、内存空间、数据通信与传递、Cache访问、负载均衡以及循环分块和向量化等方法进行了讨论。通过在内存数据库领域优化性能的应用,对使用三层优化方法发挥MIC众核技术优势进行了分析与展望。

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 three- level 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.