Energy and Thermal Aware Buffer Cache Replacement Algorithm · 2010. 5. 8. · 2 Research Summary...

MSST 2010

Jianhui Yue, Yifeng Zhu,

Zhao Cai, and Lin Lin

Electrical and Computer Engineering

University of Maine

Energy and Thermal Aware Buffer Cache

Replacement Algorithm

2

Research Summary

Memory power consumption

Consumes 50% more power than CPU on server

Consumes 2 times more power than disks in a specific

supercomputing center

Power dissipation increases cooling cost

Power and thermal aware buffer cache replacement algorithm

Limit replacements on a smaller set of memory ranks without hurting

hit rate

Save up to 27% energy

Reduce temperature 5.45 ℃ than LRU

3

Outline

Background of memory systems

New buffer cache replacement algorithm

Evaluation

Conclusion and future work

4

Memory Chip Organization

DIMM includes one/two ranks which is an independent pwr. mgt. unit

5

DRAM Power States

Read

Precharge

Powerdown

Write

144 mW

396 mW405 mW

12.6 mW

10 cycles2 cycles

DDR2 533MHz 512Mb device

Inactive/Active

Background Power

I/O Power

• Each rank includes multiple DDRx devices and provides data to data bus in union

• The rank is basic memory power management unit

6

DMA Overlapping

Multiple DMA transfers on different

PIC buses to access the same

memory rank simultaneously in a time

multiplex way.

Multiplexing multiple slow DMA

transmissions to one memory rank can

reduce active background power

Disk DMA

Channels

Memory Chips

Disk Array

Network DMA

Channels

Network

I/O Requests

Network

Adapters

Network I/O

DMA

Disk I/O

DMAA1

B1 A2B2

B3

A3

B0

A0

7

Fully-Buffered DIMM(FBDIMM)

Memory Controller

AMB

AMB

AMB

southbound linknorthbound link

write data +cmdread data

DIMM 0

DIMM 1

DIMM n

n is up to 7

1. Increases memory

density, reliability and

speed Includes read data

link and write data link

2. AMB acts as pass-

through switch

3. AMB consumes power

8

Outline

Background of memory systems


Evaluation


9

Bottom Most Replacement Algorithm (BMA)

Main Idea

Limit replacements to a smaller number of memory

ranks without hurting hit rate much.

Method

Choose the victim rank with the most number of cold

blocks in the bottom of LRU stack

Find victim block from the specified victim rank rather

than LRU block for energy saving

Update victim rank for adaption to blocks’ recency

changes

10

Choose Victim Rank

D

E

G

H

I

J

A

C

B

rank0 rank1 rank2

E

D

A

I

G

B

J

H

C

LRU stack

bottom

top Memory Physical Layout

LRU

Request

block

Victim

block

Victim

rankD

E

G

H

I

J

A

C

B

rank0 rank1 rank2

E

D

A

I

G

B

J

H

C

LRU stack

bottom


Bottom Most Algorithm (BMA)

Request

block

Victim

block

Victim

rank

Window = 5

victim_rank = 0last_victim_cnt = 3

11

Access Sequence: W, X, Y, Z

D

E

G

H

I

J

A

C

B

rank0 rank1 rank2

E

D

A

I

G

B

J

H

C

LRU stack

bottom


LRU

Request

block

Victim

block

Victim

rankD

E

G

H

I

J

A

C

B

rank0 rank1 rank2

E

D

A

I

G

B

J

H

C

LRU stack

bottom


Request

block

Victim

block

Victim

rank

Window = 5


W A 0 W 0A


12


E

G

H

I

J

W

B

D

C

rank0 rank1 rank2

E

D

W

I

G

B

J

H

C

LRU stack

bottom


LRU

Request

block

Victim

block

Victim

rank

W A 0

rank0 rank1 rank2

E

D

W

I

G

B

J

H

C

LRU stack

bottom


Request

block

Victim

block

Victim

rank

W A 0

Window = 5



E

G

H

I

J

W

B

D

C

X B 1 X D 0

13


rank0 rank1 rank2

E

D

W

I

G

X

J

H

C

LRU stack

bottom


LRU

Request

block

Victim

block

Victim

rank

W A 0

X B 1

rank0 rank1 rank2

E

X

W

I

G

B

J

H

C

LRU stack

bottom


Request

block

Victim

block

Victim

rank

W A 0

X D 0

Window = 5



G

H

I

J

W

X

C

E

D

G

H

I

J

W

X

B

E

C

Y C 2 Y E 0

14

After Y swapped in

rank0 rank1 rank2

E

D

W

I

G

X

J

H

Y

LRU stack

bottom


LRU

Request

block

Victim

block

Victim

rank

W A 0

X B 1

Y C 2

rank0 rank1 rank2

Y

X

W

I

G

B

J

H

C

LRU stack

bottom


Request

block

Victim

block

Victim

rank

W A 0

X D 0

Y E 0

Window = 5



H

I

J

W

X

Y

D

G

E

H

I

J

W

X

Y

B

G

C

15

Choose New Victim Rank

rank0 rank1 rank2

E

D

W

I

G

X

J

H

Y

LRU stack

bottom


LRU

Request

block

Victim

block

Victim

rank

W A 0

X B 1

Y C 2

rank0 rank1 rank2

Y

X

W

I

G

B

J

H

C

LRU stack

bottom


Request

block

Victim

block

Victim

rank

W A 0

X D 0

Y E 0

Window = 5



H

I

J

W

X

Y

D

G

E

H

I

J

W

X

Y

B

G

C

16


rank0 rank1 rank2

E

D

W

I

G

X

J

H

Y

LRU stack

bottom


LRU

Request

block

Victim

block

Victim

rank

W A 0

X B 1

Y C 2

rank0 rank1 rank2

Y

X

W

I

G

B

J

H

C

LRU stack

bottom


Request

block

Victim

block

Victim

rank

W A 0

X D 0

Y E 0

Window = 5



H

I

J

W

X

Y

D

G

E

H

I

J

W

X

Y

B

G

C

Z D 0 Z 1B

17

Outline

Background


Evaluation


18

Simulator and I/O Trace Characteristics

19

TPC-C Results

active

↓idle

↑

AMB

DIMM

20

LM-TBE Results

idle

↑

active

↓

AMB

DIMM

21

MSN-BEFS Results

idle

↑

active

↓

AMB

DIMM

22

Outline

Background


Evaluation


23

Conclusion and Future Work

The first method to save buffer cache memory energy without

data migration

Limit replacement to smaller set of memory rank without

sacrificing hit rate for memory

Saves up to 27% energy and reduces temperature 5.45 ℃than LRU

Implement this algorithm at Linux Kernel

Evaluate algorithm at real physical machine with I/O intensive

workload

24

Thanks !

25

DRAM Thermal

TrafficforwardingiclocalTraffstaticAMB PPPP

26

Simulation Parameters

Search window 1024 Thresh window 1024 Block size 64KB

Energy and Thermal Aware Buffer Cache Replacement Algorithm · 2010. 5. 8. · 2 Research Summary...

Documents

Transcript of Energy and Thermal Aware Buffer Cache Replacement Algorithm · 2010. 5. 8. · 2 Research Summary...