CSE 691: Energy-Efficient Computing Lecture 3 SLEEP: full-system Anshul Gandhi 1307, CS building [email protected]
Dec 23, 2015
CSE 691: Energy-Efficient ComputingLecture 3
SLEEP: full-systemAnshul Gandhi
1307, CS [email protected]
Number of servers worldwide:
Google: 1-2%
2013:Google: > 1M serversMicrosoft: ~1M serversAmazon: <1M servers
Tencent: +700K servers!
power_nap paper
US data centers: 100 billion kWh by 2011?? $$
Server utilization: <30%
Idle server power: 60% of peak
Idle periods: ~secondswhy important?
What does 30% utilization mean?
Utilization data
Existing techniques1. Consolidation
2. Sleep states
3. Throttling (DVFS)
PowerNap1. Simple idea (only 2 states)• Minimize power draw in sleep• Fast transitions
2. Model (power and response time)
3. PowerNap vs DVFS
4. RAILS
agile paper
agile• PowerNap was NOT implemented
agile took first REAL step towards that
• Static consolidation vs Dynamic consolidation
• How to minimize latency penalties of dynamic consolidation? 3 ideas.
• agile: dynamic virtualization +
PowerNap implementation
agile: main problem
agile: low-power statesTurbo
C0 P states T states
C1
C1E
C2
C3
S0 C6
S1
S2
S3
G0 S4
G2 S5
G3
agile: power vs. latency
agile: dynamic consolidation1. Host power-up2. VM migration3. Host power-down