Parallel execution for conflicting transactions

Multicore main-memory databases only obtain parallel performance when transactions do not conflict. Conflicting transactions are executed one at a time in order to ensure that they have serializable effects. Sequential execution on contended data leaves cores idle and reduces throughput. In other parallel programming contexts—not serializable transactions— techniques have been developed that can reduce contention on shared variables using percore state. This thesis asks the question, can these techniques apply to a general serializable database? This work introduces a new concurrency control technique, phase reconciliation, that uses per-core state to greatly reduce contention on popular database records for many important workloads. Phase reconciliation uses the idea of synchronized phases to amortize the cost of combining per-core data and to extract parallelism. Doppel, our phase reconciliation database, repeatedly cycles through joined and split phases. Joined phases use traditional concurrency control and allow any transaction to execute. When workload contention causes unnecessary sequential execution, Doppel switches to a split phase. During a split phase, commutative operations on popular records act on per-core state, and thus proceed in parallel on different cores. By explicitly using phases, phase reconciliation realizes two important performance benefits: First, it amortizes the potentially high costs of aggregating per-core state over many transactions. Second, it can dynamically split data or not based on observed contention, handling challenging, varying workloads. Doppel achieves higher performance because it parallelizes transactions on popular data that would be run sequentially by conventional concurrency control. Phase reconciliation helps most when there are many updates to a few popular database records. On an 80-core machine, its throughput is up to 38× higher than conventional concurrency control protocols on microbenchmarks, and up to 3× on a larger application, at the cost of increased latency for some transactions. Thesis Supervisor: Robert T. Morris Title: Professor Thesis Supervisor: Eddie Kohler Title: Associate Professor, Harvard University