Detail

Transactions on Mergeable Objects in Shared-Memory

Under high contention, serializability for transactions results in frequent aborts. This limits possible parallelism and results in performance degradation. In this talk, we introduce a new transactional semantics, Mergeable Transactions, which allows concurrent transactions on the same objects to execute in parallel. Instead of aborting and re-executing, the conflicting updates on shared objects are merged using type specific semantics.

We show that mergeable transactions outperform serializable transactions under high contention workloads.