CAVE: channel-aware buffer management scheme for solid state disk

  • Authors:

  • Sung Kyu Park;Youngwoo Park;Gyudong Shim;Kyu Ho Park

  • Affiliations:

  • Korea Advanced Institute of Science and Technology (KAIST), Guseong-dong, Yuseong-gu, Daejeon, Korea;Korea Advanced Institute of Science and Technology (KAIST), Guseong-dong, Yuseong-gu, Daejeon, Korea;Korea Advanced Institute of Science and Technology (KAIST), Guseong-dong, Yuseong-gu, Daejeon, Korea;Korea Advanced Institute of Science and Technology (KAIST), Guseong-dong, Yuseong-gu, Daejeon, Korea

  • Venue:
  • Proceedings of the 2011 ACM Symposium on Applied Computing
  • Year:
  • 2011

Quantified Score

Hi-index 0.00

Visualization

Abstract

Exploiting a multi-channel architecture of Solid State Disk (SSD) is a role of Flash Translation Layer (FTL) until now. A multi-channel FTL scheme increases I/O parallelism by spreading out pages in a logical block to multiple channels. However, this scheme has high garbage collection overhead for reclaiming invalid pages, thus resulting in the performance degradation. In order to overcome this problem, we assign a write buffer to exploit a multi-channel architecture. In this paper, we propose a novel buffer management scheme, called Channel-Aware Victim Eviction (CAVE). The key idea of the CAVE scheme is to evict multiple victims whose number is equal to the number of NAND flash memories when a write buffer is full for increasing I/O parallelism. Because a write buffer exploits a multi-channel architecture, we can use a 1-channel FTL scheme, thus reducing garbage collection overhead. We develop a trace-driven simulator for evaluating the CAVE scheme. The hit ratio and execution time are used as performance metrics. In the hit ratio, the CAVE scheme can achieve a similar result with a conventional method which uses a multi-channel FTL scheme though it evicts more victims at a time. In the execution time which consists of the write time and garbage collection time, a result shows that the CAVE scheme can reduce it by 55.5% - 97.4% in block-level LRU using SYSMARK compared to the conventional method.