Adaptive Concurrency Control: Despite the Looking Glass, One Concurrency Control Does Not Fit All

Use of transactional multicore main-memory databases is growing due to dramatic increases in memory size and CPU cores available for a single machine. To leverage these resources, recent concurrency control protocols have been proposed for main-memory databases, but are largely optimized for specific workloads. Due to shifting and unknown access patterns, workloads may change and one specific algorithm cannot dynamically fit all varied workloads. Thus, it is desirable to choose the right concurrency control protocol for a given workload. To address this issue we present adaptive concurrency control (ACC), that dynamically clusters data and chooses the optimal concurrency control protocol for each cluster. ACC addresses three key challenges: i) how to cluster data to minimize cross-cluster access and maintain load-balancing, ii) how to model workloads and perform protocol selection accordingly, and iii) how to support mixed concurrency control protocols running simultaneously. In this paper, we outline these challenges and present preliminary results.