On Making Transactional Applications Resilient to Data Corruption Faults

Multicore architectures are becoming increasingly prone to transient faults and data corruption. Relying on a multicore architecture is the common solution for increasing performance and scalability of core applications including transactional applications. In this paper we present SoftX, a low-invasive protocol for supporting execution of transactional applications relying on speculative processing and dedicated committer threads. Upon starting a transaction, SoftX forks a number of threads running the same transaction independently. The commit phase is handled by dedicated threads for optimizing synchronization´s overhead. We conduct an evaluation study showing the performance obtained with the implementation of SoftX on a 48 cores AMD machine, running List, Bank and TPC-C benchmarks. Results reveal better performance than classical replication-based fault-tolerant systems and limited overhead with respect to non fault-tolerant protocols. We ported SoftX to a message-passing architecture, Tilera TILE-Gx. Hardware message-passing is an important emerging trend in multicore architectures. Our experiments on Tilera show that SoftX is still more efficient than replication.