以MapD为代表的图分析数据库系统通过GPU、Phi等新型众核处理器来支持高性能分析处理，在面向复杂数据模式时，连接操作仍然是重要的性能瓶颈.近年来，异构处理器逐渐成为高性能计算的主流平台，内存连接性能的研究从多核CPU平台扩展到新兴的众核处理器，但众多的研究成果并未系统地揭示连接算法性能、连接数据集大小、硬件架构之间的内在联系，难以为未来异构处理器平台的数据库提供连接平台优化选择策略.以面向多核CPU、Xeon Phi、GPU处理器平台的内存连接优化技术为目标，通过优化内存哈希表设计，实现以向量映射替代哈希映射操作，消除哈希代价对内存连接算法的影响，从而更加准确地测量内存连接算法在多核CPU的cache大小、Xeon Phi的cache大小、Xeon Phi的并发多线程、GPU的SIMT （单指令多线程）机制等硬件相关因素影响下的性能特征.实验结果表明，缓存与并发多线程机制是提高内存连接算法性能的重要影响因素.缓存机制对于满足cache大小的连接操作具有性能优势，而GPU的并发多线程机制则在较大表的连接操作中具有较高的性能，Xeon Phi则在满足其L2 cache大小的连接操作中具有最高性能.实验结果揭示了内存连接操作性能与异构处理器硬件特性的联系，为未来异构处理器平台内存数据库查询优化器提供了优化策略.

Graph analysis database such as MapD employs the emerging manycore architecture GPU and Phi processors to support high performance analytical processing, where the join operation is still the performance bottleneck when facing complex data schemas. In recent years, as heterogeneous processors come to be main-stream high performance computing platforms, the researches of in-memory join performance extend the focuses from multicore to the emerging manycore platforms. However those efforts have not uncover the inner relationships among join algorithm performance, join dataset size and hardware architectures, and cannot provide sufficient join selection strategies for databases under the future heterogeneous processor platforms. This paper targets in-memory join optimization techniques on multicore, Xeon Phi and GPU processor platforms. By optimizing hash table design, this work uses vector mapping instead of hash mapping to eliminate the hashing overhead effects for performance, so that the in-memory join performance characteristics influenced can be measured by hardware factors such as multicore cache size, Xeon Phi cache size, Xeon Phi simultaneous multi-threading mechanism, and GPU SIMT (single instruction multiple threads) mechanism. The experimental results show that caching and simultaneous massive-threading mechanism are key factors to improve in-memory join algorithm performance. Caching mechanism performs well for cache fit join operations, the simultaneous massive-threading mechanism of GPU does well for big table joins, and Xeon Phi achieves the highest performance for L2 cache fit joins. The experimental results also exploit the relationship between in-memory join performance and heterogeneous processor hardware features, and provide optimization policy for in-memory database query optimizer on future heterogeneous processor platforms.

TP311

国家自然科学基金（61732014，61772533）；国家高技术研究发展计划（863）（2015AA015307）；中央高校基本科研业务费专项资金（16XNLQ02）