SCMCFS: A Complex File System For Massive SCM

When using NVM devices, the I/O performance gap narrows between main memory and external storage devices. Researchers present storage class memory(SCM) to build a universal high-speed storage system in computer. Due to the limitations of integration of NVM and the number of DIMM slot, DIMM and PCIe interface NVM devices should coexist in the massive SCM. But two types of NVM devices have many differences such as physical interface, access performance and storage capacity. Moreover, there are also different features between files, metadata and partial data in a file. However, the existing main memory and file system management algorithms are only suited for single storage mediums, which cannot be directly applied in massive SCM mixing with two types of NVM devices. In this paper, A file system named SCMCFS is introduced, aiming at implementing massive SCM with hybrid NVM devices.Firstly, we analyze the characteristics of the two types of NVM devices and the drawbacks of traditional file system. In order to manage two types of NVM device uniformly and efficiently in massive SCM, we introduce the architecture of SCMCFS.Secondly, we give the designs of the metadata organization and structure in SCMCFS. The head-tail layout and space management based on two layer radix-tree are provided to unify logic space between two type NVM devices and avoid the problem of capacity limitation. Next we introduce complex file structure, dynamic file data distributed strategy and asymmetric call in strategy, which build data buffer within file, to speed up the access response and improve I/O performance. Then the prototype of complex file system(named UFS) is implemented based on Fuse and evaluated by Postmark. The result verifies that UFS, with complex file structure, not only improves about from 30% to 63% I/O performance compared with Ext3 on PMBD, but also avoids the capacity limitation of DIMM interface NVM devices.Thirdly, The consistent mechanism is introduced in SCMCFS to maintain file system reliability. We utilize DIMM interface NVM devices as journaling devices to improve efficiency and flexibility of consistency. Next we design hybrid consistent mechanism, which can dynamically adjust consistent mechanism by the size of updating, to provide strong consistency guarantees and reduce the extra time overhead for SCMCFS. Then we implement the prototype of three consistent mechanisms base on PMFS for analyzing, which are clflush/mfence, tornbit RAWL and two-fence mechanism. The result proves that clflush/mfence and tornbit RAWL mechanism outperform two-fence mechanism 10% when the size of updating are small, while the extra time overhead of two-fence mechanism reduces from 8% to 170% than the other two mechanisms when having large-size update.Finally, the prototype of SCMCFS is implemented in Linux, and we evaluate its performance through multiple benchmarks, including Fio, Postmark and Filebench with common dataset. Compared to Ext2 and Ext4 on PMBD, SCMCFS improves sequential read speed 4.4x and 5x, sequential write speed 2.8x and 1.9x, IOPS 45% and 62%, while obtains the similar I/O performance with PMFS. In file system consistency, compared to Ext4, SCMCFS reduces the execution time of one byte consistent-update by 60%~90%, and reduces the total overhead of consistency by 50%~92%.