CDAC: Content-Driven Deduplication-Aware Storage Cache Yujuan Tan , - - PowerPoint PPT Presentation

▶
cdac content driven deduplication aware storage cache
SMART_READER_LITE
LIVE PREVIEW

CDAC: Content-Driven Deduplication-Aware Storage Cache Yujuan Tan , - - PowerPoint PPT Presentation

Dec 30, 2022 439 likes •867 views

CDAC: Content-Driven Deduplication-Aware Storage Cache Yujuan Tan , Jing Xie, Congcong Xu, Zhichao Yan, Hong Jiang, Yajun Zhao, Min Fu, Xianzhang Chen, Duo Liu, Wen Xia Outline Introduction and Motivation Design of CDAC Performance

slide-1
SLIDE 1

CDAC: Content-Driven Deduplication-Aware Storage Cache

Yujuan Tan, Jing Xie, Congcong Xu, Zhichao Yan, Hong Jiang, Yajun Zhao, Min Fu, Xianzhang Chen, Duo Liu, Wen Xia

slide-2
SLIDE 2

Outline

 Introduction and Motivation  Design of CDAC  Performance Evaluation  Conclusion

slide-3
SLIDE 3

Outline

 Introduction and Motivation  Design of CDAC  Performance Evaluation  Conclusion

slide-4
SLIDE 4

Cache Deduplication

Identifying and removing redundant data to reduce data footprint

  • The deduplication overhead, would degrade the

performance of the overall storage system

  • Be carefully designed and managed to reap the benefits
  • f increased logical capacity and cache hit ratios
slide-5
SLIDE 5

CacheDedup

Source Addresses Index Fingerprints Store Cache Blocks

Metadata Cache Data Cache

Stores deduplicated unique data blocks Stores the source addresses and data fingerprints of these blocks

  • D-LRU, D-ARC are designed based on this architecture[1]
  • Data Cache and Metadata Cache are separately managed and

accessed

[1] LI, W., JEAN-BAPTISE, G., RIVEROS, J., NARASIMHAN, G.,ZHANG, T., AND ZHAO, M. CacheDedup: In-line Deduplication for Flash Caching. In USENIX FAST’16 (Feb. 2016).

slide-6
SLIDE 6

Analysis of D-ARC and D-LRU

  • For 4KB, D-ARC and D-LRU is 6.91% and

11.85% higher than ARC and LRU on average

  • For 8KB, 16KB and 32KB, D-ARC are

5.38%, 3.00% and 2.65% higher than ARC on average, and D-LRU are 8.70%, 5.58% and 3.77% higher than LRU on average

  • For 4KB, the read hit ratios of D-LRU and

D-ARC are 76.5% and 89.2% of that of OPT

  • For 8KB, 16KB and 32KB, the read hit

ratios of D-LRU are only 31%, 12.3% and 6% of that of OPT, and D-ARC’s read hit ratios are only 82.9%, 73.9% and 49% of that of OPT

(Cache size:40%)

  • Read hit ratio from WebVm
slide-7
SLIDE 7

Analysis of D-ARC and D-LRU

  • Overall hit ratio from WebVM

As the block size increases, the benefits of deduplication become limited and their hit ratios decreases significantly

slide-8
SLIDE 8

Existing algorithm analysis

Analysis

  • Do Not fully utilize the

characteristics of deduplication

——Missed the opportunity to effectively leverage the intensity of content redundancy and sharing. ——Based on this discovery, we proposed RCE technology

  • Cache space utilization is also low

——Read/write alignment causes a large amount of invalid data in the cache. ——Based on this discovery, we proposed BHI technology

slide-9
SLIDE 9

Outline

 Introduction and Motivation  Design of CDAC  Performance Evaluation  Conclusion

slide-10
SLIDE 10

CDAC Design

  • Architecture
  • Based on CacheDedup architecture

——Data Cache stores the data blocks, Metadata Cache stores the source addresses and data fingerprints of these blocks ——The source address and its corresponding data block do not need to be fetched in or evicted out synchronously

  • Key technology
  • Reference Count Eviction (RCE)

——Focuses on exploiting the blocks’ content redundancy and their intensity of content sharing among source addresses

  • Bitmap Hotness Identification (BHI)

——More accurately identify the content hotness of the block, especially for large blocks, minimizing false-positive hot block identifications

  • Terminology
  • Free block

—— The block in Data Cache that there is no source address in Metadata cache pointing to it

  • Reference count

——The total number of the source addresses pointing to that block

CDAC focus on how to select the source addresses to be deleted from Metadata Cache to generate the free blocks to improve the cache hit ratios

slide-11
SLIDE 11

How to choose, B or C?

Referenced-Count based Eviction（RCE）

A1 B1 C1 A B C A2 B2 B3 Metadata cache Data cache LRU Reference count 2 3 1 MRU:the position of most recently used LRU:the position of least recently used MRU

  • The higher the reference count, the more requests will be associated

with this data block, and so the hotter the data block wil be

RCE takes both reference counts and access locality into consideration

  • However, using reference counting as the only hint to find the block

to be replaced is not sufficiently effective

slide-12
SLIDE 12

Referenced-Count based Eviction（RCE）

  • RCE only focuses on the source addresses in the LRU position
  • RCE divides the data blocks into two categories
  • Category 1: the data block that is referenced only once
  • ——No other source address is associated with it, RCE will

delete it

  • Category 2: the block that is referenced multiple times
  • ——Move the source address to the MRU position to keep it
  • ——Further observe how the reference count changes in the

next cycle

  • ——If the rate of decline exceeds the threshold in the next

cycle, the source address is deleted

  • A cycle: The time required for the source address to go from the MRU

position to the LRU position

slide-13
SLIDE 13

Referenced-Count based Eviction（RCE）

A1 B1 C1 A B C A2 B2 B3 Metadata cache Data cache LRU Reference count 2 3 1 MRU:the position of most recently used LRU:the position of least recently used MRU

  • Access Order:

——B3, B2, C1, A2, B1, A1, D1, B3, E1, F1……

slide-14
SLIDE 14

Referenced-Count based Eviction（RCE）

A1 B1 C1 A B C A2 B2 Metadata cache Data cache LRU Reference count 2 2 1 MRU:the position of most recently used LRU:the position of least recently used MRU B3

  • Access Order:

——B3, B2, C1, A2, B1, A1, D1, B3, E1, F1……

slide-15
SLIDE 15

Referenced-Count based Eviction（RCE）

A1 B1 C1 A B C A2 B2 Metadata cache Data cache LRU Reference count 2 2 1 MRU:the position of most recently used LRU:the position of least recently used MRU B3

  • Access Order:

——B3, B2, C1, A2, B1, A1, D1, B3, E1, F1……

slide-16
SLIDE 16

Referenced-Count based Eviction（RCE）

A1 B1 C1 A B C A2 Metadata cache Data cache LRU Reference count 2 1 1 MRU:the position of most recently used LRU:the position of least recently used MRU B3 B2

  • Access Order:

——B3, B2, C1, A2, B1, A1, D1, B3, E1, F1……

slide-17
SLIDE 17

Referenced-Count based Eviction（RCE）

A1 B1 C1 A B C A2 Metadata cache Data cache LRU Reference count 2 1 1 MRU:the position of most recently used LRU:the position of least recently used MRU B3 B2

  • Access Order:

——B3, B2, C1, A2, B1, A1, D1, B3, E1, F1……

slide-18
SLIDE 18

Referenced-Count based Eviction（RCE）

A1 B1 A B A2 Metadata cache Data cache LRU Reference count 1 2 1 MRU:the position of most recently used LRU:the position of least recently used MRU B3 B2 D1 D

  • Access Order:

——B3, B2, C1, A2, B1, A1, D1, B3, E1, F1……

slide-19
SLIDE 19

Referenced-Count based Eviction（RCE）

A1 B1 A B A2 Metadata cache Data cache LRU Reference count 1 2 2 MRU:the position of most recently used LRU:the position of least recently used MRU B2 D1 D B3

  • Access Order:

——B3, B2, C1, A2, B1, A1, D1, B3, E1, F1……

slide-20
SLIDE 20

Referenced-Count based Eviction（RCE）

A1 B1 A B A2 Metadata cache Data cache LRU Reference count 1 2 2 MRU:the position of most recently used LRU:the position of least recently used MRU B2 D1 D B3

  • Access Order:

——B3, B2, C1, A2, B1, A1, D1, B3, E1, F1……

slide-21
SLIDE 21

Referenced-Count based Eviction（RCE）

A1 B1 A B A2 Metadata cache Data cache LRU Reference count 1 1 1 MRU:the position of most recently used LRU:the position of least recently used MRU B2 D1 D B3

  • Access Order:

——B3, B2, C1, A2, B1, A1, D1, B3, E1, F1……

slide-22
SLIDE 22

Referenced-Count based Eviction（RCE）

A1 B1 A B A2 Metadata cache Data cache LRU Reference count 1 1 1 MRU:the position of most recently used LRU:the position of least recently used MRU B2 D1 D B3

  • Access Order:

——B3, B2, C1, A2, B1, A1, D1, B3, E1, F1……

slide-23
SLIDE 23

Referenced-Count based Eviction（RCE）

  • Access Order:

——B3, B2, C1, A2, B1, A1, D1, B3, E1, F1……

Reference count B1 B Metadata cache Data cache LRU 1 1 1 MRU:the position of most recently used LRU:the position of least recently used MRU B2 D1 D B3 E E1 A2

slide-24
SLIDE 24

Referenced-Count based Eviction（RCE）

  • Access Order:

——B3, B2, C1, A2, B1, A1, D1, B3, E1, F1……

Reference count B1 B Metadata cache Data cache LRU 1 1 1 MRU:the position of most recently used LRU:the position of least recently used MRU B2 D1 D B3 E E1 A2

slide-25
SLIDE 25

Referenced-Count based Eviction（RCE）

  • Access Order:

——B3, B2, C1, A2, B1, A1, D1, B3, E1, F1……

B1 B Metadata cache Data cache LRU Reference count 1 1 1 MRU:the position of most recently used LRU:the position of least recently used MRU D1 D B3 E E1 A2

slide-26
SLIDE 26

Bitmap based Hotness Identification（BHI）

  • Read and write alignment causes the waste of much cache space

——The hotness/coldness recognition of traditional algorithms has ignored the valid content of each cached block for each access.

  • BHI identifies hot/cold blocks based on finer-grained access

——breaks a block into multiple small parts and then uses bitmaps to record the access status of each part

represents for the accessed part represents for the non-accessed part 1：the block has been accessed in the current cycle 0：the block has not been accessed in the current cycle

  • Flag: used to identify the source addresses that have not been accessed for

a long time, even if their access parts greatly exceed the threshold

  • The current cycle: The time required for the source address to go from the

MRU position to the LRU position

slide-27
SLIDE 27

Bitmap based Hotness Identification（BHI）

T1 A(1,2,3) Access Time Access Order The block stored in cache T2 B(3) T3 C(1,2,3,4) T4 D(4) A(2,3)

1

A

1

B

1

A

1

C

1

B A

T5

1

slide-28
SLIDE 28

Bitmap based Hotness Identification（BHI）

T1 A(1,2,3) Access Time Access Order The block stored in cache T2 B(3) T3 C(1,2,3,4) T4 D(4) A(2,3)

1

A

1

B

1

A

1

C

1

B A

1

C B A

Step1:

T5

1 1

slide-29
SLIDE 29

Bitmap based Hotness Identification（BHI）

T1 A(1,2,3) Access Time Access Order The block stored in cache T2 B(3) T3 C(1,2,3,4) T4 D(4) A(2,3)

1

A

1

B

1

A

1

C

1

B A

1

C B A

Step1:

1

C A

T5

Step2:

1 1

slide-30
SLIDE 30

Bitmap based Hotness Identification（BHI）

1

D

T1 A(1,2,3) Access Time Access Order The block stored in cache T2 B(3) T3 C(1,2,3,4) T4 D(4) A(2,3)

1

A

1

B

1

A

1

C

1

B A A

1

C

1

C B A

Step1:

1

C A

T5

Step2: Step3:

1 1

B3 hits in the cac he. B3 hits in the cac he.

slide-31
SLIDE 31

Bitmap based Hotness Identification（BHI）

1

D

T1 A(1,2,3) Access Time Access Order The block stored in cache T2 B(3) T3 C(1,2,3,4) T4 D(4) A(2,3)

1

A

1

B

1

A

1

C

1

B A A

1

C

1

A

1

C

1

C B A

Step1:

1

C A

T5

Step2: Step3:

1 1 1

D

slide-32
SLIDE 32

CDAC Design

  • Combining RCE and BHI together
  • CDAC first uses BHI to check if the source address in the LRU

position is recognized as a cold source address The combination of BHI and RCE enables CDAC to more accurately identify the cold blocks and associated addresses to improve the cache hit ratios

  • If it is, CDAC uses RCE to identify if it needs to be deleted
  • The source address in the LRU position needs to be constantly

checked and deleted until a free block is found

slide-33
SLIDE 33

Outline

 Introduction and Motivation  Design of CDAC  Performance Evaluation  Conclusion

slide-34
SLIDE 34

Performance Evaluation

  • CDAC implementations: CDAC-ARC, CDAC-LRU
  • Baselines: D-ARC, D-LRU, ARC, LRU
  • Baseline approaches
  • Parameter settings
  • Cache size: From 20% to 80% of the working set size
  • Block size: From 4KB to 64KB
  • Decline rate threshold for reference count in RCE: 50%
  • The threshold in BHI : 50%
  • The size of the smallest part of each block in BHI: 4KB
  • Evaluation metric
  • Cache hit ratio
slide-35
SLIDE 35

Performance Evaluation

Name Total I/Os (GB) Working Set(GB) Write to read ratio Unique Data(GB) WebVM 54.5 2.1 3.6 23.4 Homes 67.3 5.9 31.5 44.4 Mail 1741 57.1 8.1 171.3

  • Trace statistics

These traces were collected from a VM hosting the departmental websites for webmail and online course management (WebVM), a file server used by a research group (Homes), and a departmental mail server (Mail) [1].

[1] KOLLER, R., AND RANGASWAMI, R. I/o deduplication: Utilizing content similarity to improve i/o performance. In Usenix Conference on File & Storage Technologies (2010).

slide-36
SLIDE 36

Performance Evaluation

  • Hit Ratios (4KB sized block)

Using the reference counts that represent the intensity of the content sharing among source addresses helps preserve a lot of hot data blocks and associated source addresses，thereby increasing the cache hit rate

Read hit ratio with 4KB block size Overall hit ratio with 4KB block size

——Contain only the performance results of RCE —— the threshold for Decline rate threshold for reference count in RCE：50%

  • Analysis
slide-37
SLIDE 37

Performance Evaluation

  • Read hit ratio

CDAC-LRU and CDAC-ARC outperform D-LRU and D-ARC by 1.95X on average

slide-38
SLIDE 38

Performance Evaluation

  • Overall hit ratios

Both CDAC-LRU and CDAC-ARC obtain higher read hit ratios and overall hit ratios than their corresponding baseline approaches

slide-39
SLIDE 39

Performance Evaluation

  • Analysis of the experimental results
  • CDAC’s performance improvement in read hit ratios is

greater than overall hit ratios

  • As the block size increases, the amount of cache space

required for maximum improvement in CDAC increases

  • CDAC’s overall performance advantages over the

baselines become more pronounced as the block size increases

slide-40
SLIDE 40

Outline

 Introduction and Motivation  Design of CDAC  Performance Evaluation  Conclusion

slide-41
SLIDE 41

Conclusion and Acknowledgement

 Analyzing the existing deduplication-aware caching algorithms  Proposing CDAC based on CacheDeduparchitecture, which focuses on

exploiting the blocks’ content redundancy and their intensity of content sharing among source addresses

 Our extensive experimental results show that CDAC-LRU and CDAC-

ARC outperform D-LRU and D-ARC by 1.95X on average under real- world traces

 Acknowledgement: We are very grateful to Professor Ming Zhao

from Arizona State University for providing us with CacheDedup Prototype and many instructive comments.

slide-42
SLIDE 42

Thanks! Q&A