ZNSwap: un-Block your Swap

Abstract

We introduce ZNSwap , a novel swap subsystem optimized for the recent Zoned Namespace (ZNS) SSDs. ZNSwap leverages ZNS’s explicit control over data management on the drive and introduces a space-efficient host-side Garbage Collector (GC) for swap storage co-designed with the OS swap logic. ZNSwap enables cross-layer optimizations, such as direct access to the in-kernel swap usage statistics by the GC to enable fine-grain swap storage management, and correct accounting of the GC bandwidth usage in the OS resource isolation mechanisms to improve performance isolation in multi-tenant environments. We evaluate ZNSwap using standard Linux swap benchmarks and two production key-value stores. ZNSwap shows significant performance benefits over the Linux swap on traditional SSDs, such as stable throughput for different memory access patterns, and 10 × lower 99th percentile latency and 5 × higher throughput for memcached key-value store under realistic usage scenarios.

ZenFS+: Nurturing Performance and Isolation to ZenFS

“…To minimize the garbage collection overhead, the authors selectively put sstables that have similar lifetimes into the same zone, relaxing the write and space amplification from garbage collection. Shai et al presented an approach to using ZNS SSD as a swap device [49]. Because ZNS SSD has a little cost of firmware-level garbage collection, the swap performance has significantly increased.…”

Identity of university Chinese heritage language learners in Hong Kong

“…The Zoned Namespace (ZNS) interface [14,19] is a new NVMe Command Set [4] that exposes flash-based SSD internals to the host for host-level storage management. Figure 1 shows the architecture of a ZNS SSD based on the Western Digital Ultrastar DC ZN540 model, which we use in this paper and also previous studies (e.g., [18,42,65]). The ZNS interface abstracts a ZNS SSD as append-only zones, each of which has a maximum size called the zone capacity.…”

“…Other studies focus on the performance and management aspects of ZNS SSDs. Examples include improving the performance of log-structured merge-tree stores [59] for ZNS SSDs [37,47,48], optimizing host-level garbage collection [20,26,30,65], proposing new ZNS interfaces for efficient zone management [30,55,58], enabling ZNS SSDs for swap storage [18], designing new I/O scheduling for improved intra-zone parallelism [17], extending Zone Append for sub-block data appends [62], and ensuring crash consistency on F2FS backed by ZNS SSDs [46]. While the above studies focus on a single ZNS SSD, RAIZN [42] exposes a ZNS SSD array as a single ZNS interface to applications, and focuses on fault tolerance, correctness, and crash consistency.…”

LearnedFTL: A Learning-based Page-level FTL for Improving Random Reads in Flash-based SSDs

“…When building a learned model over this GTD entry, it is necessary for the GC process to collect the valid pages across multiple flash blocks, and these blocks also contain data (PPNs) belonging to multiple GTD entries. As a result, the GC process needs a To address this issue, we propose a group-based allocation strategy in our LearnedFTL to reduce the GC overhead and simplify the model training [5]. The basic idea is to divide GTD into groups of consecutive entries, referred to as GTD entry group.…”

“…Figure 10: The principle of virtual PPN translation.large amount of data movement, which significantly increases the complexity and overhead of the model training process.To address this issue, we propose a group-based allocation strategy in our LearnedFTL to reduce the GC overhead and simplify the model training[5]. The basic idea is to divide GTD into groups of consecutive entries, referred to as GTD entry group.…”