Designing And Optmizing Of NUMA-aware PCM File System

The emerging storage class memory(SCM), such as Phase Change Memory(PCM), which has both characteristics of disk and memory in storage capacity and access speed, respectively, provides physical support for constructing data-centred computing architecture. Currently, researches based on SCM mainly focus on managing storage space and extending service life. NUMA(Non-Uniform Memory Access) architecture has currently become the mainstream server architecture, however, that how to effictively deploy SCM on NUMA system has not received sufficient concern. As the characteristic of NUMA architecture, there must exist substantial new and challenging problems in deploying SCM on NUMA system.In this paper, we first analyze the characteristics of NUMA, which is a main architecture used in commercial server currently, then we analyze the Performance Monitor Unit(PMU) of Intel Xeon E7 processor which bases on NUMA architecture and also analyze memory access data flow on it. As processor uncore plays an important role in memory access performance, we monitor the change trend of the counter of performance events relating memory access under different memory access pressure in three circumstances, which are LLC contention, processor remote memory access, and memory controller contention. We found that those performance event counters can reflect the memory access pressure of nodes on NUMA system. Known this, we propose a memory access pressure assessing model that uses a machine learning algorithm to assess the memory access pressure of nodes on NUMA system based on performance event counters relating to memory access of processor uncore. Our memory access pressure assessing model can rightly predict the least memory access pressure node in the NUMA system as high as 63%.As for the storage capacity of PCM, we design a PCM file system based on NUMA architecture. According to the characteristic of NUMA system, we found that access speed of PCM from local node is higher than that from remote node when access pressure of local PCM is low, and access speed of PCM from remote node is higher than that from local node when access pressure of local PCM is relative high after profiling the performance of PCM file system using file system performance benchmark. Based on those experiment results, we implement a NUMA-aware PCM file system which integrate the memory access pressure assessing model into PCM file system. After file accessing request sent by process, PCM file system assesses memory access pressure of each node on NUMA system, then it decides whether migrate the process to the prediction node which may obtain the higher PCM access speed. Experiment results show that our PCM file system has great NUMA awareness ability in most scenarios after the test of two file system performance profiling benchmark, Iozone and Postmark.