Silo: Speculative Hardware Logging for Atomic Durability in Persistent Memory

Persistent memory (PM) provides data persistency. Due to this property, guaranteeing atomic durability becomes important for applications running on PM in order to ensure the crash consistency for a group of updates. To this end, hardware logging has received many attentions by overlapping the log operations and transaction execution. Unfortunately, existing approaches regard logs as backups, which inevitably increases the log writes to PM, thus exacerbating the limited endurance of PM and imposing constraints on the write ordering.

This paper proposes Silo, a speculative hardware logging design to ensure atomic durability with ultra-low overheads. Unlike existing studies, Silo exploits a speculative methodology and regards logs as data to make the common case fast. In practice, system crashes or power failures rarely occur for a machine. Hence, we do not need to write logs to back up data in most of the running time. Based on this observation, Silo temporarily maintains the undo+redo logs on chip during transaction execution. After the transaction commits, Silo leverages the new data in these on-chip logs to in-place update the PM data region, instead of conservatively writing logs as useless backups in common cases where no crash occurs. In this way, Silo significantly reduces the write overheads. If a crash occurs, Silo still efficiently flushes these on-chip logs to PM for recovery without any loss of correctness. Experimental results demonstrate that Silo significantly improves the transaction throughput by 4.3×, and reduces the memory writes by 76.5% compared with state-of-the-art designs.