File systems and applications try to implement their own update protocols to guarantee system-wide and application-level consistency, respectively, since ensuring data consistency is one of the most crucial aspects of computing systems. For example, in order to guarantee system-wide consistency of file system data, the journaling file system is widely adopted in modern computing systems, ranging from battery-powered smartphones to enterprise servers. Moreover, in order to guarantee application-level consistency of critical data, database engines operate in either Rollback or Write-Ahead Logging (WAL) mode, while supporting the atomicity of transaction execution. However, we found that the storage devices are substantially under-utilized when preserving system-wide and application-level consistency, since it generates massive storage write traffic with many disk cache flush operations and force-unit-access (FUA) commands.
The focus of this dissertation is to provide both high performance and strong consistency in the file system-level without any modifications on applications, device drivers, and storage firmware. For this, we first introduce a novel journaling scheme, called Delta Journaling (DJ), to reduce journaling overhead in mobile devices by using a small-sized NVRAM, such as phase-change memory (PCM) or spin-transfer torque magnetic RAM (STT-MRAM), efficiently. DJ uses two separate journal areas, one in the NAND flash storage and another in the NVRAM, and stores a journal block as a compressed delta in the small-sized NVRAM only when the compressed delta is small enough. Our experimental results show that DJ enhances transaction throughput by up to 2.23 times for a real-world workload, with only 4 MB NVRAM. Second, in order to cover a wide range of storage devices in computing systems, we introduce a novel journaling file system, called Delta-FS. Delta-FS makes three technical contributions to provide both high performance and strong consistency for applications. Firstly, in order to remove all storage accesses with disk cache flush operations and FUA commands that are generated upon periodic commit operations or frequent fsync() calls, Delta-FS adopts a full journal mode of Ext4 and uses a small-sized NVRAM for a file system journal. Secondly, in order to reduce access latency and space requirements of NVRAM journal, Delta-FS attempts to journal compressed differences of modified blocks, rather than full blocks themselves, while efficiently exploiting the byte-addressable characteristics of NVRAM. Lastly, in order to relieve explicit checkpointing overhead, Delta-FS aggressively reflects the checkpoint transactions to the file system area in the unit of specified region, by exploiting the idle time of storage device. Experimental results clearly show that Delta-FS outperforms Ext4 by up to 64.2 times with only 128 MB NVRAM, by reducing the journal write traffic along with the number of disk cache flush operations and FUA commands dramatically via our novel optimization schemes.