TO FLUSH or NOT: Zero padding in the file system with SSD devices

Abstract

Guaranteeing data persistency is extremely critical in the file system, and for data consistency, most file systems use the FLUSH command to force data to the persistent storage media, such as SSD devices. However, the FLUSH command significantly decreases the overall performance for flushing the data and the L2P mapping table stored on the write buffer. In this paper, we propose a novel alternative scheme, called zero padding, that ensures the same effect as physically issuing the FLUSH command while mitigating the performance problem of the FLUSH command. To achieve this, zero padding fully utilizes both the behaviors of the write buffer and the bandwidth of SATA 3.0 interface. We have successfully realized zero padding in ext4 and JBD2 with a small modification, called ext4-zp, and experimentally compared it with the traditional ext4 journal modes. Our experimental results show that ext4-zp provides the best throughput in various workloads and outperforms ext4 with ordered mode by up to 40%.