BUDW: Energy-Efficient Parallel Storage Systems with Write-Buffer Disks

Energy-Efficient Parallel Storage Systems with Write-Buffer Disks

Xiaojun Ruan and Xiao Qin

Computer Science and Software EngineeringSamuel Ginn College of Engineering

Auburn University

My Research Group: 2011

Xiaojun Ruan

3

Overview of the Project

Performance of Secure Disk Systems

[IEEE NAS09]

Energy Efficiency

Security Solid State Drives

Design, Model, Simulate, And Evaluate

Disk Systems with Buffer Disks

[ACM SAC09][ICPP09]

Energy-EfficientDistributed StorageSystems [IPCCC10]

Enhancing Internal Parallelism of SSDs[To Be Submitted11]

BUD

Message Passing Interface with

Enhanced Security[IPCCC 2010]

Energy-Efficient Dynamic Voltage Scaling[ICCCN07]

04/10/2023 4

Annual Data Center Electricity Usage and Electricity Price increase Every year

Electricity Usage in Data Centers

04/10/2023 5

• The average power consumption of TOP10 supercomputing systems is 1.32 Mwatt.

Storage

37%

Server

40%

Network 23%

Dell’s Texas Data Center

Energy Efficiency of Supercomputers

Electrical Cost of Data Centers

Using 2010 Trends Scenario◦ Server and Data Centers Consume 110 Billion kWh

per year◦ Assume average commercial end user is charged 9.46

kWh◦ Disk systems can account for 27% of the energy cost

of data centers

04/10/2023 6

Server and data centers may have an electrical cost of 10.4 billion dollars.

04/10/2023 7

Energy Consumption of Disks

8

Active State: high energy consumption

Power States of Disks

Active

StandbyState transition penalty

Standby State: low energy consumption

9

A10000RPM Hard Drive may take 10.9 seconds to wake up!

A Hard Disk Drive

Parallel Disks

Performance

Energy Efficiency

Challanges

Performance Oriented:

• Best Performance

• Huge Electricity Bills

Energy Efficiency Oriented:

• Worst Performance?

• Small Electricity Bills

12

Basic Idea of BUD

• Keep Disks in Standby mode as long as possible

• Reduce Status Transitions as many as possible

IBM Ultrastar 36Z15

04/10/2023 13

Transfer Rate 55 MB/s Spin Down Time: TD 1.5 s

Active Power: PA 13.5 W Spin Up Time: TU 10.9 s

Idle Power: PI 10.2 W Spin Down Energy: ED 13 J

Standby Power: PA 2.5 W Spin Up Energy: EU 135 J

Break-Even Time: TBE 15.2 S

A Parallel Disk System with a Write Buffer Disk

15

The BUD Architecture

Data Disks can serve requests without buffer disks when workload is high

Auburn University 16

Sum of Requests in Buffer (SRB)

• SRB is Number of the buffered requests targeting at the same data disk.

• SRB is set by administrators• Once SRB is satisfied, spin up the targeted

data disk, dump all those data, then spin the disk down.

Scheduling Strategy

DynAmic Request Allocatio

n algorithm for Writes

17

Put the Request in the Buffer Disk Queue

Yes

Data Disk Availabe?

No

Write the Request into a Buffer Disk

Is theTargeted Data Disk Availabe?

No

Write the Request into Data Disk

Yes

Yes

New Request?

No

To buffer enough requests targeting at the same data disk

Example

Buffer Disk

Requests Queue

18

Auburn University Xiaojun Ruan 19

From Design to Simulation

20

Simulation Environment

System Parameter. IBM 36Z15 UltraStar IBM 40GNX Travel Star

Rotations Per Minute 10000 RPM 5400 RPM

Working Power 13.5 W 3 W

Standby Power 2.5 W 0.25 W

Spin up Energy 135 Joule 8.7 Joule

Spin down Energy 13 Joule 0.4 Joule

Spin up Time 10.9 sec 3.5 sec

Spin Down Time 1.5 sec 0.5 sec

Transfer Rate 52.8 MB/s 25 MB/s

Auburn University 21

Workloads

Impact of SRB—Low Workload, UltraStar

22

Auburn University Xiaojun Ruan 23

Non-Buffer Experiments

Auburn University 24

BUD with IBM 40GNX TravalStar

Buffer Disk Number and Workload-- UltraStar

25

Auburn University 26

Energy Consumption

E = Active Energy Consumption + Standby Energy Consumption + Transition Penalty

Auburn University Xiaojun Ruan 27

From Simulation to Real Implementation

28

An Energy-Efficient Cluster Storage System

29

Implementation (no buffer disks)

30

Implementation (with buffer-disks)

31

Experimental Design

• Disk Category I/O Node 1

• Data Disk 1: WesternDigital 400, 20GB• Data Disk 2: WesternDigital 400, 20GB

• Disk Category I/O Node 2

• Data Disk 1: WesternDigital 400, 20GB• Data Disk 2: Maxtor D740X-6L, 20GB

32

Experimental Design

• Disk Category I/O Node 1• Buffer Disk: Maxtor DiamondMax Plus 9• Data Disk 1: WesternDigital 400, 20GB• Data Disk 2: WesternDigital 400, 20GB

• Disk Category I/O Node 2• Buffer Disk: Seagate Barracuda 7200• Data Disk 1: WesternDigital 400, 20GB• Data Disk 2: Maxtor D740X-6L, 20GB

33

34idle time gap is 200sidle time gap is 100s

04/10/2023 35

GreenFS[ACM EuroSys 2008]

Massive Arrays of Idle Disks[SC 2002]

Popular Data Concentration[ACM ICS 2004]

Previous Research

Download the presentation slideshttp://www.slideshare.net/xqin74

Google: slideshare Xiao Qin

‹#›

http://www.eng.auburn.edu/~xqin

My webpagehttp://www.eng.auburn.edu/~xqin

Download Slides at slideshare

http://www.slideshare.net/xqin74

Auburn University 40

Questions?