TaP: Tablebased prefetching for storage caches

1
TaP Table-based prefetching for storage caches

• Mingju Li, Elizabeth Varki, Swapnil Bhatia
• University of New Hampshire
• Arif Merchant
• Hewlett-Packard Labs

2
Main contribution of TaP

• Effective sequential stream detection
• Efficient prefetch cache size adaptation

3
Introduction

• Our interpretation of sequential prefetching
• Motivation for TaP
• What is TaP
• How TaP works
• TaPs evaluation
• Future work

4
Sequential prefetching interpretation

• Three modules of a sequential prefetching
  technique
• Classifications of current storage sequential
  prefetching techniques

5
Three modules of a sequential prefetching
technique

• Sequential detection module
• Search for sequentiality in I/O workload
• Determines the effectiveness of prefetching
• Prefetching module
• Prefetching degree
• Prefetching trigger point
• Cache management module
• Cache segment (prefetch cache )
• Cached read data classification
• Cache size management
• Cache replacement policy

6
Classification of sequential prefetching
techniques

• Prefetch on Miss (PoM)
• Prefetch on Hit (PoH)
• CaP Cache-based prefetching
• TaP Table-based prefetching (new)
• Always Prefetch (AP)

7
Classification of sequential prefetching
techniques
8
Motivation for sequential prefetching

• Powerful controller and large cache
• Principle of locality
• Low temporal locality
• Cache hierarchy
• Possible spatial locality
• Sequential data layout on disks

9
Motivation for TaPsequential detection module

• Low temporal locality from higher level caching
• Low sequential locality in the workload from
  multiple stream interleaving

• No knowledge about file system or applications

10
Motivation for TaPsequential detection module

• Current techniques
• AP and PoM has no detection module
• PoH Has inefficient detection module
• CaP Cache-based sequential detection
• (cache pollution)
• (Suffer from short history)

11
Motivation for TaPcache management module

• Most techniques mix the prefetched data with
  other data in the cache
• Keep prefetched data in the cache after hit
• Fixed cache segment size not efficient cache
  usage

12
What is TaP

• A sequential prefetching technique
• For storage read caches only
• Effective sequential detection module
• Efficient (prefetch) cache management module
• Simple to implement

13
How TaP works

• Prefetch cache
• Dedicated cache portion for prefetched data
• TaP table
• Sequential detection
• Prefetch cache size adaptation

Table entry
Request address
cache full flag
14
TaP table for sequential detection

• Tracks the request addresses in FIFO list
• if request i is a cache hit
• Serve i from cache, prefetch i1
• If i is a cache miss
• if i is a table hit serve i from disk,
  prefetch i 1 into cache
• else i is a table miss serve i from disk,
  insert i 1 into table

11 (false)
86 (false)
35 (false)
12(data)
13(data)
10
11
85
34
45
12
46 (false)
workload
Prefetch cache
TaP table
15
Advantage of TaP sequential detection

• More workload history is available for detection
• High useful prefetch ratio
• Fraction of prefetches that are hit

16
TaP table for prefetch cache sizing

• Increasing
• Cache full, evicted request address inserted into
  table
• Flag in TaP table for evicted request

26 (false)
12(True)
24 (false)
12(data)
34(data)
24 (false)
12(True)
68 (false)
34(data)
35(data)
68 (false)
46 (false)
27(data)
35(data)
26
23
67
45
12
46 (false)
13(data)
workload
Prefetch cache
TaP table

• Decreasing periodically

17
TaPs evaluation

• Simulation tool
• DiskSim a validated, accurate storage system
  simulator
• Sequential detection evaluation
• Prefetch cache sizing evaluation

18
Sequential detection evaluation SPC2-like
workload (hit rate)
19
Sequential detection evaluation SPC2-like
workload (response time)
20
Sequential detection evaluation SPC2-like
workload (useful prefetch ratio)
21
Sequential detection evaluationsequential
random workload (mean hit rate)
22
Sequential detection evaluationsequential
random workload (response time)
23
TaPs Prefetch cache sizing evaluation

• Given a small prefetch cache, TaP always gets
  better performance than other algorithms compared
• TaP uses smaller prefetch cache size to get the
  same cache hit rate as CaP

24
Cache size comparisonCaP and TaP
25
Cache sizing evaluation sequential random
workload (mean hit rate)
26
Cache sizing evaluation sequential random
workload (response time)
27
TaP prefetch cache size adaptation visualization
28
Conclusion

• Sequential stream detection and prefetching is
  important for storage only caching
• Prefetched data should be treated differently
  from other cached data
• The TaP table is the key component of TaP for
• Effective sequential stream detection
• Efficient prefetch cache usage

29
Future work

• Cache sizing
• Table sizing
• Prefetching module
• Re-reference hit

30
Questions ?
Thank you!