A Thorough Analysis of Page Fault Handling in Persistent Memory Systems

The new technologies for building persistent devices have reached a point where these devices can be added to the processor bus and accessed via regular load/store instructions. Commonly know as Persistent Memory (PM), these devices have renewed the research interest in systems used to program them. One important implementation technique used by a class of these systems is the use of DRAM as shadow memory, which allows the use of contemporary hardware transactions. However, these systems have an important drawback: if DRAM is considerably smaller than PM, the performance can be degraded due to excessive paging. Despite this, few previous works have looked into that issue. We provide in this paper, for the first time, a thorough analysis of the performance of PM systems when the amount of DRAM is smaller than that of PM. We also present a user-level page handling mechanism that can be integrated in any current PM system. Whereas previous works have considered only synthetic workloads, our study uses a realistic benchmark. The experimental evaluation shows that the final performance under paging is heavily influenced by how often the transactions enters the Single Global Lock (SGL) mode, that is, the amount of conflicts caused by the paging mechanism.