WARCIP: W : Write A Ampli lific ication Reduction b by C Clus - - PowerPoint PPT Presentation

WARCIP: W : Write A Ampli lific ication Reduction b by C Clus ustering I I/O Pages

Jing Yang, Shuyi Pei and Qing Yang

• Dept. of Electrical, Computer, and Biomedical Engineering, University of Rhode Island

Kingston, RI, USA 02881 Shenzhen DAPU Microelectronics Co., Ltd Shenzhen, China {jyang,spei,qyang}@ele.uri.edu

Garbage Collection Overheads

• Drags down I/O performance
• Consumes a significant amount of back-end bandwidth of SSDs
• Reduces SSD’s endurance
• Valid page migration causes extra writes to NAND flashes

Valid Invalid Invalid Valid Valid Valid free free GC candidate block GC appending point (open block) Valid page migration Different Lifetime

Garbage Collection Overheads (cont.)

WA can be reduced by minimizing lifetime variance of pages in a flash block.

Lifetime

>

Time Block 1 Block 2 Block 1 Block 2 GC Two valid pages need to be moved out No page needs to be moved

a b

Existing Solutions

• Traditional SSD only has one or two

appending points

• Pages with various lifetimes are mixed

in the same erase unit

• Multi-stream SSD has multiple

appending points

• Data from different streams are

physically separated to different blocks

• Less lifetime variance in a flash block
• Reduce write amplification

App 1 App 2 App n Open Block

…

Write request SSD App 1 App 2 App n Open Block 2

…

Open Block 1 Open Block 3 Open Block n

…

SSD Mixed in the same erase unit Separate in different erase unit

Challenges in Stream Management

VM 1 VM 2 VM 30

… …

s1 s2s3s4s5 s6s7 s8

OS

Multi-stream SSD

Limited stream number Changing workloads Require application or system modification

Challenges in stream management (cont.)

VM 1 VM 2 VM 30

… …

s1 s2s3s4s5 s6s7 s8

OS

Multi-stream SSD Multi-stream SSD

O B O B O B O B O B Treat SSD as a black box

WARCIP: Write Amplification Reduction by Clustering I/O Pages

• Minimize rewrite interval (RWI) variance of pages in a flash block
• Learn current I/O patterns to fine-tune its clustering activities
• Device-side data clustering solution

Greedy Clustering

Open Block 1 Open Block 2 Open Block 3 Open Block 8

…

Page A Timestamp Table Page A RWI OB1 Avg RWI OB2 Avg RWI OB3 Avg RWI OB4 Avg RWI OB5 Avg RWI OB6 Avg RWI OB7 Avg RWI OB8 Avg RWI OB3 Avg RWI

Dynamic Split-and-Merge

1 2 3

Open Blocks

Split

1 5 3 2

Open Blocks

1 2 3

Open Blocks

Merge

1 3

Open Blocks

Feedback

• Collect valid pages’ information during GC
• Double the recorded rewrite interval of each valid page

Page a Page b

a b

Timestamp Table

SSD Write Amplification Improvement

SSD Erasure Count Reduction

SSD Throughput

Fewer page migrations release backend bandwidth Changed write pattern causes more WA Self-adaptive to new pattern

Conclus usions ns

• We proposed WARCIP, a new approach to minimize the negative

impact of garbage collection on SSD

• WARCIP groups pages into a block according to their rewrite intervals to

reduce write amplification caused by page migrations

• Self-optimizing and adaptive to new I/O patterns
• Future work
• Wear leveling