Paper Info
Title
An I/O scheduler based on fine-grained access patterns to improve SSD performance and lifespan
Abstract
Although the many benefits delivered by Solid State Disks (SSDs), they also pose some unique and serious challenges to I/O and file system designers. Unlike HDDs and other memory devices, SSDs cannot perform in-place updates. A block has to be erased before it can be re-written. Moreover, the costs of different SSD operations are highly asymmetric. A write operation in an SSD is an order of magnitude slower than a read operation, and an erase operation is in turn an order of magnitude slower than a write. Moreover, a block can endure only a limited number of erasures before it wears out. Most SSDs employ a log-structured Flash-Translation-Layer (FTL) to solve the not-in-place update problem. The unique operations of the FTLs, together with the asymmetric overheads of different operations, imply that many traditional solutions optimized for HDDs do not work well for SSDs. For example, sequential writes that are not perfectly aligned to the flash block boundary, may reduce performance and increase wearing overhead. In this paper, we proposed a novel I/O scheduler which is based on fine-grained access patterns in a per-process per-stream manner. These patterns are used to guide a set of novel scheduling policies, including pre-alignment, inner-padding, write merging, merging-and-splitting, to improve the write performance of SSDs that adopt log-structured FTLs. Simulation results show that these policies can improve write performance by up to 60%. Moreover, the schemes reduce SSD erasure cycle by up to 64%, which is directly translated to a major improvement on the lifespan of SSDs.
Year
DOI
Venue
2014
10.1145/2554850.2554971
SAC
Keywords
Field
DocType
design,experimentation,measurement,input/output,sequencing and scheduling,performance
File system,Computer science,Scheduling (computing),Parallel computing,Input/output,Merge (version control),Flash storage,Solid-state,Erasure,Overhead (business)
Conference
Citations 
PageRank 
References 
1
0.35
12
Authors
2
Name
Order
Citations
PageRank
Mingyang Wang121.08
Yiming Hu263944.91