HiperTM: High Performance, Fault-Tolerant Transactional Memory

Abstract

We present HiperTM, a high performance active replication protocol for fault-tolerant distributed transactional memory. The active replication paradigm allows transactions to execute locally, costing them only a single network communication step during transaction execution. Shared objects are replicated across all sites, avoiding remote object accesses. Replica consistency is ensured by a OS-Paxos, an optimistic atomic broadcast layer that total-orders transactional requests, and b SCC, a local multi-version concurrency control protocol that enforces a commit order equivalent to transactions' delivery order. SCC executes write transactions serially without incurring any synchronization overhead, and runs read-only transactions in parallel to write transactions with non-blocking execution and abort-freedom. Our implementation reveals that HiperTM guarantees 0% of out-of-order optimistic deliveries and performance up to 1.2× better than atomic broadcast-based competitor PaxosSTM.