Cutting power consumption in

cloud
l d computing
ti

阳振坤
Senior scientist, Baidu
yangzhenkun@baidu com
yangzhenkun@baidu.com
2009.6.26
 Agenda
• Background information
• Power consumption in IDC
• Cutting
C tting po
power
er cons
consumption
mption
• Summary
 What is cloud computing
• An emerging computing technology that uses the
internet and central remote servers (IDC) to
maintain data and applications
• Enabling much more efficient computing by
centralizing storage
storage, memory
memory, processing and
bandwidth
¾ Machine idle ratio: ~70% vs
vs. 90% in traditional
• A kind of large-scale distributed (usually
heterogeneo s) ssystem
heterogeneous) stem
 IDC: some numbers
• The data center in Dallas, Oregon: ~50
50 MW
¾ 50MW*0.8/200W=0.2M
¾ Average electricity consumption in USA:
~900kwh/month/family, or 1.25KW
• Power consumption is the major cost and constraint
of IDC
• About 7000 IDCs in USA
 IDC’s
IDC s infrastructure

~AC-1
AC 1

~AC-2
AC 2

• Usually it is critical to protect the

cooling
li system b
by UPS
 Power consumption in IDC
• Machines
¾ 50,000 machines (300W each)
consume 1.5 MW by
th
themselves
l
• Network device
• Cooling system
• UPS array and their batteries
¾ For both machines and cooling
system
• Lighting
Li hti
• Power Usage Effectiveness
(PUE)
 Cooling system
• Water
Water-cooling
cooling vs. air-cooling
air cooling
¾ Heat transfer coefficient: 100:1
¾ Air-cooling
Air cooling device is simpler and cheaper and more
suitable for home and ordinary office
¾ Thee vast
as majority
ajo y o
of machines
ac es a are
eaair-cooling
coo g
• “Free cooling”
 UPS array
• Contribute to about 11 percent of power
consumption and also produce heat
• Obsolete UPS array?
¾ Internal battery as standby power for machines
¾ Activating standby power (usu
(usu. diesel generators) for
cooling system in a few minutes after power failure
 Lighting
• Contribute to about 3 percent of power
consumption and also produce heat
• To cut power consumption by lighting
¾ Energy-saving lamps
¾ Voice activated switch
Voice-activated switch…
 Machines
• Processor
• Memory
¾ Larger capacity vs
vs. smaller capacity
¾ FB/ECC vs. ordinary RAM
• Power
P supply
l efficiency
ffi i
¾ ATX: 70%~75%
¾ High (400) volt DC vs. AC
• Hard disk
¾ HDD vs. SSD disk
¾ 3.5” vs. 2.5” hard disk
 Machines
• Redundant components
¾ Graphics adapter, USB ports, DVD driver, sound chip…
• Cooling system
¾ Water-cooling vs. air-cooling
• Machine
M hi idlidle ratio
ti
¾ Usually 70% or lower. Can it be cut idle ratio more?
• Standby or hibernation
¾ Is it possible in cloud computing?
 Data replication in cloud computing
• Data replication、heartbeat,
p ,…
¾ Some data become unavailable if many machines unavailable

0 2 0 2
5 4 5
1 3
0 8
D 5 7 7
3
M
1 2 3
8 1
6
4 6 4 6
7 8
 Cluster: an example
• Requirements
¾ 50TB of data, 2,000,000 queries/s and 90% cache hit
rate
¾ Single machine: 100~400 queries/s, 0.15~1.8TB disk
capacity
• Cluster planning
¾ 2,000,000 (1 90%)/400 = 500 machines
2,000,000*(1-90%)/400
¾ 50TB*3/500 = 0.3TB/machine
• Can we make half or 1/3 of machines standby or
hibernated?
 Cluster: an example (cont’d)
(cont d)
• Make machines standby/hibernated
y one by
y one
¾ May lead to mass data shuffle, or
¾ Requires data distribution
0 2 0 2
5 4 5
1 3
0 8
D 5 7 7
3
M
1 2 3
8 1
6
4 6 4 6
7 8
 Summary
• Power is the major constraint and cost of IDC
• To cut power consumption
¾ Computer machines
¾ Cooling system
¾ UPS
¾ System infrastructure…
• Whole
Wh l iindustry
d t should
h ld b
be iinvolved
l d

• Q&A