Shingled Magnetic Recording (SMR) is a new technique for increasing areal data density in hard drives. Drive-managed SMR (DM-SMR) drives employ a shingled translation layer to mask internal data management and support block interface to the host software. Two major challenges of designing an efficient shingled translation layer for DM-SMR drives are metadata overhead and garbage collection overhead. In this paper we introduce SMaRT, an approach to Shingled Magnetic Recording Translation which adapts its data management scheme as the drive utilization changes. SMaRT uses a hybrid update strategy which performs in-place update for the qualified tracks and out-of-place updates for the unqualified tracks. Background Garbage Collection (GC) operations and on-demand GC operations are used when the free space becomes too fragmented. SMaRT also has a specially crafted space allocation and track migration scheme that supports automatic cold data progression to minimize GC overhead in the long term. We implement SMaRT and compare it with a regular Hard Disk Drive (HDD) and a simulated Seagate DM-SMR drive. The experiments with several block I/O traces demonstrate that SMaRT performs better than the Seagate drive and even provides comparable performance as regular HDDs when drive space usage is below a certain threshold.
|Original language||English (US)|
|Title of host publication||Proceedings of the 15th USENIX Conference on File and Storage Technologies, FAST 2017|
|Number of pages||13|
|State||Published - 2019|
|Event||15th USENIX Conference on File and Storage Technologies, FAST 2017 - Santa Clara, United States|
Duration: Feb 27 2017 → Mar 2 2017
|Name||Proceedings of the 15th USENIX Conference on File and Storage Technologies, FAST 2017|
|Conference||15th USENIX Conference on File and Storage Technologies, FAST 2017|
|Period||2/27/17 → 3/2/17|
Bibliographical noteFunding Information:
We thank the anonymous reviewers and our shepherd Erik Riedel for their insightful comments on earlier drafts of the work. This work was partially supported by NSF awards 130523, 1439622, and 1525617.