Our paper investigating the impact of fast, non-volatile memories on high-performance applications is available. We used our Moneta storage array and some emulation tricks to understand that performance impact of attaching fast non-volatile memories (like PCM) to both the DDR and PCIe busses. The results show that while these technologies can deliver enormous gains in raw performance, the application-level gains lag significantly. The data in the paper argue that the cause of this disparity is software inefficiency (e.g., scheduling overheads that help in disk-based systems, but just add latency for very fast storage).
Adrian will be presenting the paper at Supercomputing in New Orleans this November.