Peak Power Aware Parallel Hash Join

With the development of SSD, how to fully profit from the new storage has become a hot research topic in recent years. Hash join is a core query operator and its performance optimization and power control are two critical issues in data center nowadays. This thesis focused on the performance optimization and power control of hash join on SSDs.The contributions of the thesis are as follows:1. Performance features of late SSDs were summarized. This paper derived some algorithmdesign principles by looking into the relationship between access granularity and queuedepth under read and write mixture workloads.2. A parallel hash join algorithm on SSDs and its optimized buffer allocation scheme wereproposed. By utilizing internal parallelism of SSDs, Grace Hash Join was parallelizedaccordingly and a new buffer allocation scheme was invented.3. The peak power characteristics of parallel hash join were analyzed. We predicted poweraccording to core utilization and discussed the relationship among buffer size, I/Othroughput and core peak power.4. Hash join power controlling methods were proposed by restricting the number of coresactivated.Experiment results demonstrated that our parallel hash join could fully take advantage of the SSD performance gain, and the buffer allocation scheme worked well without wasting memory resource. The power analysis results matched well with the real power profiling while the power controlling methods are efficient and works well.