ZAB: Fast and Efficient Write Buffer Algorithm for Optimizing Parallelism inside Zoned Namespace Storage Device

초록

Today, smart TVs play an increasingly important role in creating a smart home environment. With the intensifying storage demands of applications, smart TVs have been adopting NAND-based storage devices (i.e., eMMCs, UFSs, SD cards, and SSDs) as their primary storage media. In this paper, we introduce emerging zoned namespace storage devices (e.g., ZUFS and ZNS SSDs) that can be adopted in smart TVs owing to their desirable features: isolated performance and internal parallelism at the hardware level. We also focus on the write buffer of the zoned namespace storage device (i.e., ZNS SSD) and then propose a new write buffer algorithm for emerging zoned namespace storage devices, called ZAB. The key idea of ZAB is to efficiently isolate noisy from normal neighbors and use the write buffer space to maximize the benefits of internal parallelism. To do this, we introduce three novel designs: (1) a noisy zone classifier to efficiently isolate noisy neighbors in multiple zones, (2) flash-level granularity to group data logically in the buffer for good intra-chip parallelism, and (3) fairness-oriented scheduling to further enhance built-in parallelism using inter-chip parallelism. We implemented a prototype of ZAB based on ConfZNS and compared it with a comprehensive option by varying the write buffer algorithms. The experimental results indicate that ZAB improves overall performance by up to 7 and 4.3 times that of the baseline in real-world and multi-tenant workloads, respectively. © 2014 IEEE.

키워드