SLA: A Cache Algorithm for SSD-SMR Storage System with Minimum RMWs

Zheng, Xuda; Zhang, Chi; Duan, Keqiang; Wu, Weiguo; Yan, Jie

doi:10.1007/978-3-030-95391-1_37

Cited by 2 publications

(1 citation statement)

References 20 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Buffer blocks must be periodically written back to accommodate the incoming blocks. In general, the persistent buffer adopts a zone-level write-back mechanism to evict buffer blocks within the same zone that reduces the write amplification generated by cleanup activities [6,7,19,[32][33][34][35]. As shown in Figure 4, the RMW operation of the cleanup process consists of three steps: reading valid blocks from the native zone, collecting victim blocks, and writing back the merged blocks.…”

Section: Motivationmentioning

confidence: 99%

Amphisbaena: A Novel Persistent Buffer Management Strategy to Improve SMR Disk Performance

Zhang,

Yu,

Nie

et al. 2024

Applied Sciences

Self Cite

View full text Add to dashboard Cite

The explosive growth of massive data makes shingled magnetic recording (SMR) disks a promising candidate for balancing capacity and cost. SMR disks are typically configured with a persistent buffer to reduce the read–modify–write (RMW) overhead introduced by non-sequential writes. Traditional SMR zones-based persistent buffers are subject to sequential-write constraints, and frequent cleanups cause disk performance degradation. Conventional magnetic recording (CMR) zones with in-place update capabilities enable less frequent cleanups and are gradually being used to construct persistent buffers in certain SMR disks. However, existing CMR zones-based persistent buffer designs fail to accurately capture hot blocks with long update periods and are limited by an inflexible data layout, resulting in inefficient cleanups. To address the above issues, we propose a strategy called Amphisbaena. First, a two-phase data block classification method is proposed to capture frequently updated blocks. Then, a locality-aware buffer space management scheme is developed to dynamically manage blocks with different update frequencies. Finally, a latency-sensitive garbage collection policy based on the above is designed to mitigate the impact of cleanup on user requests. Experimental results show that Amphisbaena reduces latency by an average of 29.9% and the number of RMWs by an average of 37% compared to current state-of-the-art strategies.

show abstract