CRAZNS: A Case for Conventional Namespace Support for RAID with ZNS SSDs

Abstract

Zoned Namespace (ZNS) SSDs are flash-based SSDs that maintain several zones in storage. Each zone has its own write pointer to prevent any write requests from occurring in front of or behind it. While ZNS SSDs achieve improved write performance with the write pointer, they also face limitations as in-place updates are not allowed. This limitation poses a challenge in building a Redundant Array of ZNS SSDs as metadata and partial parity logs can be done more efficiently with overwrites. In this paper, we advocate for the use of a conventional namespace in ZNS RAID, that is, RAID that uses ZNS SSDs, and to support this, we design and implement CRAZNS, a ZNS RAID-5 that makes use of a conventional namespace. Compared to RAIZN, the state-of-the-art ZNS RAID, CRAZNS uses 4 GB extra storage space for the conventional namespace, but is able to use the maximum number of zoned namespaces that are possible and saves almost 26 GiB of storage space by eliminating the need for Metadata zones. Performance evaluations show that for individual applications performance between RAIZN and CRAZNS were similar, but in terms of small write throughput, CRAZNS is 1.2× higher than RAIZN. Also, CRAZNS enhances overall throughput by 1.1× over RAIZN as more zones can be kept open.