Accelerating Scan Transactions on Scan-intensive Workloads with Node Locking

抄録

Transaction processing is widely used in various applications, including credit card payments and social data analysis. To deal with these applications, scalable concurrency control protocols have been proposed. Modern techniques mainly target SEARCH and UPDATE operations and YCSB and TPC-C benchmarks are used for their evaluation. Unfortunately, the issues of improving the performance of SCAN, which is mandatory for large-scale data analysis, and avoiding phantom anomalies have not been addressed in such techniques. In this paper, we propose a novel concurrency control protocol for the efficient execution of transactional wide-range SCAN based on the optimistic concurrency control protocol. The proposed protocol, Node-locking, acquires locks of the leaf nodes in the tree index during SCAN to prevent phantoms in the read phase. The proposed protocol avoids deadlocks due to locking by applying the WaitDie protocol to index nodes. On a workload containing SCAN transactions reading 10,000 records, Silo with the proposed protocol shows up to 743 times performance improvement over the vanilla Silo protocol.