Graceful Operation of DiskDrives under Thermal Emergencies

1
Graceful Operation of Disk-Drives under Thermal
Emergencies

• Y Kim, J. Choi, S. Gurumurthi, A.
  Sivasubramaniam
• January 4, 2007
• Dept. of Computer Science and Engineering
• The Pennsylvania State University
• Dept. of Computer Science
• University of Virginia

2
Thermal Problems of Data Centers
3
Thermal Impact on Disk-Drive Reliability

• Heat-Related Problems
• Data corruption
• High off-track errors
• Head-crashes

• Thermal Emergency
• HVAC/fan break-down
• Hot neighbor components

• Possible Solutions
• Powerful cooling systems
• Duplicated over-cooling systems
• Thermal Management of Disk-Drive

• D. Anderson et al, More than an Interface
  SCSI vs. ATA, in FAST 2003.

4
Conventional Disk-Drive under Thermal Emergency
T E M P E R A T U R E
Thermal Emergency Line
No available service due to disk off !
TIME
5
Multi-speed Disk-Drive ISCA03
RPM
TIME
Power ( Platters)(RPM)2.8(Diameter)4.6
6
Power Source of Disk-Drive
7
Using Multi-Speed Disk-Drive under Thermal
Emergency
T E M P E R A T U R E
Thermal Emergency Line
Service still available for the requests under
thermal emergency !
TIME
8
Design Issue of Multi-Speed Disk-Drive under
Thermal Emergency

• Goals of thermal management with multi-speed
  disk-drive?
• Avoid thermal emergency
• Maximize performance by continuously servicing
  the requests

• Design issue of efficient multi-speed disk-drive?
• How long should it run at low speed?
• Can it service the requests at low speed?
• MDISKsimple spinning at multi-speeds while
  servicing only at one speed
• MDISKopt spinning and servicing at different
  speeds

• Control policies of multi-speed disk-drive?
• Time-based policy
• Watermark-based policy

9
Time-based Policy for MDISKsimple
T E M P E R A T U R E
Thermal Emergency Line

• Service is not available during a pre-defined
  delay.
• Long delay still causes performance-degradation
  !

TIME
10
Watermark-based Policy for MDISKopt
T E M P E R A T U R E
Thermal Emergency Line

• Service available under thermal emergency as
  opposed to MDISKsimple.
• Performance highly depends on Tlow !

TIME
11
Experiments

• We used a Performance-Thermal Simulator for
  Storage Systems (called STEAM).

Time Stamps
DiskSim (Event-Driven)
Thermal Model (Time-Step based)

• DiskSim2.0
• Performance simulator for conventional disk-drive
• Thermal Model
• To generate temperature distribution of
  disk-drive by using finite difference method to
  calculate heat flow

12
Workloads

• Scenarios for thermal emergency with different
  real I/O traces

Workloads
Initial Tamb(C)
Increased Tamb(C)
Start (s)
End (s)
Disk ()
Thermal Emergency
Simulated Time(s)
HPL Openmail
29 (C)
42 (C)
500
2,500
3,607
8
OLTP Application
29 (C)
33 (C)
5,000
30,000
43,712
24
Search Engine
29 (C)
33 (C)
2,000
12,000
15,395
6
TPC-C
29 (C)
33 (C)
2,000
10,000
15,851
4
13
Time-based policy for MDISKsimple

• HPL-Openmail

• Large RPM transition time degrades the
  performance.
• More cooling unit times prevent
  performance-degradation.

14
Watermark-based Policy for MDISKopt

• TPC-C

• RPM transitions are crucial in performance
• Low Tlow gives sufficient cooling effect on
  disk-drive to leverage the overhead of RPM
  transition time.

15
Conclusions

• We proposed a multi-speed disk-drive approach for
  thermal management of disk-drive under thermal
  emergency.
• It avoids thermal emergency while maintaining
  service availability under thermal emergency.
• RPM transition time in a multi-speed disk-drive
  should be minimized for better application.

16
Thank you !
17
Additional Slides
18
Growth in Drive Performance
Source Hitachi GST Technology Overview Charts,
http//www.hitachigst.com/hdd/technolo/overview/s
toragetechchart.html
19
Thermal Profiles (DRPM)
(a) HPL Openmail
(b) OLTP
(c) Search-Engine
(d) TPC-C
20
Control Policies for DRPM
1. Time-based Policy T temperature P
pre-defined time T read_temp() X
read_time() while(1) if( T gt
Tthreshold ) drop_rpm()
// wait for predefined time do
X read_time() while( X lt
P) increase_rpm()
else service()