DULO: An Effective Buffer Cache DULO: An Effective Buffer Cache Management Scheme to Exploit Both Management Scheme to Exploit Both

Temporal and Spatial LocalitiesTemporal and Spatial Localities

Song Jiang, Wayne State University

Feng Chen and Xiaoning Ding, Ohio State

Reported by wen

2

““Disk Wall”Disk Wall” is a Critical Issueis a Critical Issue

Disk performance for workloads without dominant sequential accesses can be seriously degraded.

Throughput of accesses to sequentially placed disk blocks can be an order of magnitude higher than that of accesses to randomly placed blocks.

Unfortunately, spatial locality of cached blocks is largely ignored and only temporal locality is considered in system buffer cache management.

3

Sequential Locality is Unique in DisksSequential Locality is Unique in Disks

Sequential Locality: disk accesses in sequence fastest

Disk speed is limited by mechanical constraints.

seek/rotation (high latency and power consumption)

OS can guess sequential disk-layout, but not always right.

4

Our solution: DULOOur solution: DULO

Exploiting DUal LOcalities (DULO) Temporal locality of program execution

Sequential locality of disk accesses

minimizing random disk accesses

Application independent approach

putting disk access information on OS map

5

Traditional efforts to break the disk bottleneckTraditional efforts to break the disk bottleneck

1. Reduce disk accesses through memory caching

By using replacement algorithms to exploit the temporal locality, and then reduce disk activities of requested blocks.

The object only to reduct block miss ignores to utilize sequential stored blocks on single disk track.

2. I/O request scheduling

Try to minimal seeks and thereafter maximal global disk throughput. e.g. SSTF, CSCAN

6

Traditional efforts... cont.Traditional efforts... cont.

3. Prefetching

prefeching manager predicts the future request patterns to fetch data in advance.

Conclusion:

I/O scheduling and prefetching are positioned at a level lower than buffer cache, so they have limited ability to cach the opportunities lost in buffer cache management.

I/O Scheduler

Disk Driver

Application I/O Requests

disk

Buffer cacheCaching &

prefetching

7

DULODULO key operationskey operations

Forming sequences

A sequence is defined as a number of blocks whose disk locations are adjacent and have been accessed during a limited time period.

sorting the sequences in the LUR stack according to their lecency and size.

sequences of large recency and size are placed close to the LRU stack bottom

8

Sequence Preview

Sequence the random disk block requested

15

DULO Design

1.Structuring LRU stack

2.Block Table: A data structure for Dual Locality

3.Forming Seqeuences

4.The DULO Replacement Algorithm

16

DULO Design: structuring LRU stack

Staging Section

Evicting Section

Correlation Buffer

Sequencing Bank

LRU Stack

The correlation buffer in DULO is similar to the correlation reference period used in the LRU-K replacement algorithm. It's size is fixed.

The sequencing bank is used to prepare a collection of blocks to be sequenced, and its size ranges from 0 to a maximum value, BANK MAX.

The blocks that are leaving the stack are sorted in the evicting section for a replacement order reflecting both their sequentiality and their recency.

17

DULO Design: Block Table Block Table a data structure for tracking disk blocksa data structure for tracking disk blocks

time1

Timestamps

time2

0

10

20

LBN: 5140 = 0*5122 + 10*512 + 20

18

9

7

10

3

8

= 9= 10

2

DULO Design:: Identifying Long Disk SequenceIdentifying Long Disk Sequence

= 7

1^

LBN : Block

N2N3

N1

N4 8

8

= 8

9

9

10

10

^

^

4^

19

7

9

10

3

82

DULO Design:: Identifying Long Disk SequenceIdentifying Long Disk Sequence

1

9

9

4

10

10

Sequence

Not a sequence

20

15

17

162

DULO Design:: Identifying Long Disk SequenceIdentifying Long Disk Sequence

1

6

17

17

Continuously Accessed

Not Continuously

Accessed

Not a Sequence (Lacking Stability)

21L=L1

DULO design: Replacement Algorithm

LRU Stack

Adapted GreedyDual Algorithm a global inflation value L , and a value H for each sequence

Calculate H values for sequences in sequencing bank:

H = L + 1 / Length( sequence )

Random blocks have larger H values

When a sequence (s) is replaced,

L = H value of s .

L increases monotonically and make future sequences have larger H values

Sequences with smaller H values are placed closer to the bottom of LRU stack

H=L0+1

L=L0

H=L0+0.25

H=L0+1

H=L0+0.25

22

Disk-Seen TASK 3: Replacement Algorithm

LRU Stack

Adapted GreedyDual Algorithm a global inflation value L , and a value H for each sequence

Calculate H values for sequences in sequencing bank:

H = L + 1 / Length( sequence )

Random blocks have larger H values

When a sequence (s) is replaced,

L = H value of s .

L increases monotonically and make future sequences have larger H values

Sequences with smaller H values are placed closer to the bottom of LRU stack

H=L1+1

H=L1+0.25

L=L1

H=L0+0.25

H=L0+1

23

DULO-Caching Principles

Moving long sequences to the bottom of stack

replace them early, get them back fast from disks

Replacement priority is set by sequence length.

Moving LRU sequences to the bottom of stack

exploiting temporal locality of data accesses

Keeping random blocks in upper level stack

hold them: expensive to get back from disks.

24

Evalution Results: extremely example

25

Evalution Results: Mixed request patterns

26

Parameter discussion

• The (maximum) sequencing bank size

• The (minimal) evicting size.

Staging Section

Evicting Section

Correlation Buffer

Sequencing Bank

LRU Stack

27

Evalution Results: bank size

An optimal bank size roughly in the range from 4MB to 16MB.

28

Bank Size

• A bank with too small size has little chance to form long sequences.

• Meanwhile, a bank size must be less than the evicting section size. This is because a large bank size causes the evicting section to be refilled too late and causes the random blocks in it to have to be evicted.

29

Evalution Results: eviction size

The larger the section size, usually means better performance. The figure does show the trend.

30

The DULO Implementation:Linux Caching

• The Linux 2.6 kernel groups all the process pages andfile pages into two LRU lists called the active list and the inactive list.

• we partition the inactive list into a staging section and an evicting section because the list is the place where new blocks are added and old blocks are replaced.

31

The DULO Implementation:Implementation Issues

• With two lists,both newly accessed pages and not recently accessed active pages demoted from the active list might be added into the inactive list and probably be sequenced in the same bank .

32

Related Work

• Algorithms such as 2Q, MQ, ARC LIRS et al, focus only on how to better utilize temporal locality, to better predict the blocks to be accessed and try to minimize page fault rate. None of these algorithms considers spatial locality

• Prefetchers cannot change the I/O request stream in any way as the buffer cache does, they can take advantage of the more sequential I/O request streams resulted from the DULO cache manager.

• Misses on the two types of blocks are equally treated without giving preference to random blocks

33

Limitations

1.DULO attempts to provide random blocks with a caching privilege. DULO can do little help when I/O requests access to random blocks that have not been accessed for a long time.

2.We adapt it to the 2Q-like Linux replacement policy. How to adapt DULO on other advanced caching algorithms and understanding the interaction between DULO replacement policy.

3. Integration between caching and prefetching polices in the DULO scheme.

34

ConclusionsConclusions

Disk performance is limited by

Non-uniform accesses: fast sequential, slow random

OS has limited knowledge of disk-layout: unable to effectively exploit sequential locality.

DULO can significantly improve I/O performance by exploiting both temporal and spatial locality.