Exploring The Green Blade Ken Lutz University of California, Berkeley LoCal Retreat, June 8, 2009.
-
date post
21-Dec-2015 -
Category
Documents
-
view
216 -
download
0
Transcript of Exploring The Green Blade Ken Lutz University of California, Berkeley LoCal Retreat, June 8, 2009.
Exploring The Green Blade
Ken LutzUniversity of California, Berkeley
LoCal Retreat, June 8, 2009
Motivation
2
• To Realize An Energy Responsive Blade and Cluster– Consolidate Load within the cluster– Turn OFF Underutilized Blades
– Explore a smaller unit of replication
– Run on Battery providing Demand Response• Platform for System Level Experiments
3
Server Power Consumption
230
15
248
87
190
13
190
13
200
14
161
19
287
48
0
50
100
150
200
250
300
350
Wat
ts
Pow
erE
dge
1850
Del
l Pow
erE
dge
1950
Sun
Fire
V60
x
Sun
Fire
x21
00 -
Cyb
er S
witc
hing
Sun
Fire
X22
00
Com
paq
DL3
60
HP
Int
egrit
y rx
2600
Server Power Consumption
Active
Idle
Maybe we can explore a new data point
ATOM Blade
• Explore a New Building Block• Inexpensive ($100) and Low Power (25W)• Based on the Intel ATOM 330 CPU
– Dual Core – 45nm– 8W MAX
• Smaller Unit of Replication so expect a cluster will have more units to do the same work
• Fine Grain Control4
Energy Responsive Node
5
Node On/OFF
AC ON/OFF
Battery Backup
UPS Controller ATOM
LoCal ATOM Cluster
• (4) ATOM Motherboards – Almost 1U– 160MB Disk Drive– 10/100 Ethernet– 2GB DRAM– USB etc– DC-DC converter with UPS switch and Battery
• Where is the power going?
6
Room for Improvement…
7
TV Encoder
Clock Generator
Audio Codec
Dc-DC ConverterBridge Chip
Intel I/O controller hub Chip
ATOM CPU
No Heat Sinks
8Not Steady State
Idle Power
9
CPU Power Control Schemes may have a limited effect
Conclusion
• The goal is to explore a new data point in cluster design
• Small nodes allow fine grained load distribution
• Ability to turn nodes off• Explore limited but distributed energy
storage• Learn about the system level problems
10