Some Joules Are More Precious Than Others:
Managing Renewable Energy in the Datacenter
Adapted from HotPower 2009
Christopher Stewart
The Ohio State University
[email protected]
with Deng Nan and Christopher Sherman
Sustainable, Green Renewable Energy
■ Benefits: Green and
Green
The grid relies on
unsustainable sources
(now)
Green House Data
□ 10,000 sq. ft. in WY
□ Wind powered
Similar datacenter plans
in CA, TX, MA, NY & NE
□ The Grid's future is
uncertain, but
sustainable sources
are a long way from
being prevalent
□ Coal still beats solar
and wind by 2X in
2030 [EIA, 2009]
Slide 2
Profitable, Green Renewable Energy
■ Benefits: Green and Green
Energy production with low
maintenance & human
costs
□ Incremental growth
AISO.net
• [Google's Solar
Panel Project]
□ Solar-powered
datatcenter in CA
• AISO's pay-go
model
□ Competitive monthly
hosting $9.95-$50
• Buy equp. or pay as
a service
□ Profitable
□ Amortize cost over
equip. lifetime
Slide 3
Profitable, Green Renewable Energy
■ Benefits: Green and Green
Energy production with low
maintenance & human
costs
□ Incremental growth
• [Google's Solar
In
2009,
7
multi-megaWatt
datacenters
AISO.net
Panel Project]
adopted
solar
panels
and/or
wind
turbines.
•
AISO's pay-go
□ Solar-powered
model
[environment
leader]
datatcenter
in CA
□ Competitive monthly
hosting $9.95-$50
□ Profitable
• Buy equp. or pay as
a service
□ Amortize cost over
equip. lifetime
Slide 4
The Intermittency Problem
Applications in the datacenter must be available 24x7
But wind and solar energy are intermittent
■ Datacenters powered by renewable energy need
backups
□ Primary options: Grid, generator, battery
□ Alternatives are either unsustainable and/or costly
■ Renewables are precious!
• renewable = joule converted from solar/wind
□ The preferred energy source
□ Available only sometimes and costly to store
Slide 5
Traditional Intermittency Solutions
■ Batteries
□ Store renewables until workload spikes
□ Loose energy in storage
□ Equipment is costly
■ Market liberalisation (net metering)
□ Sell unused renewable energy to grid
□ Sell price < Buy price
□ Limits on the amount of energy that can be sold
Slide 6
Position Statement
1. Datacenter design and management should strive to
increase the use of renewables
Opportunities

Capacity planning

Compute power should fluctuate with intermittent
outages—i.e., turn machines off
• Load balancing

Route requests to datacenters with unused
renewables


[K. Le-HotPower-2009] [A. Hopper-2008]
Migrate services to datacenters with renewables
2. Need mechanisms for renewable-aware
management
Slide 7
Intermittent Renewables are More Precious
■ Every Joule is Precious[Vahdat, 2000], right?
■ Energy efficiency alone may not be enough
□ Consider the architecture of a storage device
Policy w/ Energy Efficient
Management:
Low power
write
buffer
Power
hungry
perm.
writes
1. Put the disk to sleep
2. Fill the low power write buffer
3. Dump write buffer all at once,
in a bursty fashion
What if renewables are
available during the sleep?
A Renewable-Aware Policy
1. Put the disk to sleep
2. Fill the low power write buffer
3. Dump write buffer whenever
renewable energy is available
If write buffer fills and no
renewables are available, do
we fall back to energy
efficiency?
Slide 8
Outline
□ Studying the Impact of Intermittency
• Model of a wind-powered datacenter
• Effect on wind energy production
• Use of renewable energy in the datacenter
□ Managing Renewables at the Request Level
□ Conclusion
Slide 9
Intermittency
1.
2.
3.
4.
Datacenter Model
Wind Intermittency
Renewables in the DC
Related Work
Approach
■ Understand the flow of power through a datacenter
□ Power sources, e.g., the Grid or a renewable energy
□ Power-transfer devices, e.g., ATS, PDU
□ Power consuming equipment
■ Identify parameters that affect the use of intermittent
power from renewable sources
■ Assess the impact of these parameters
Slide 10
1. Datacenter Model
2. Wind Intermittency
3. Renewables in the DC
4. Related Work
Intermittency
Battery Backup
Power Transfer System
ATS
Grid Energy
STS/
UPS
ATS
Off-grid renewables
Primary
1 MW peak
Secondary
1 MW
Output
Power transfer system:
Input: K High voltage, turbulent AC streams
Output: N high-quality power streams to N compute devices
Slide 11
Intermittency
1. Datacenter Model
2. Wind Intermittency
3. Renewables in the DC
4. Related Work
■ Automatic transfer switch (ATS)
□ Input 2 power sources, Outputs 1 power source
□ Prevents dangerous “backfeed” between sources
• Only one source is active at a time
□ Here's how it works
1.Monitors power from the primary source
ATS
Primary
Secondary
2.When power from primary dips below K, ATS
switches (with no detectable delay) to secondary
3.When primary exhibits K power for K' seconds, ATS
switches back to primary
Output
Slide 12
Intermittency
1. Datacenter Model
2. Wind Intermittency
3. Renewables in the DC
4. Related Work
Key parameters related to the ATS
■ Power-transfer threshold
□ Power level where ATS switches to the secondary
source
□ For dependability, threshold equals peak consumption
□ For high use of renewables, set the threshold low
■ Average power consumption of the datacenter
□ When renewables are available
□ In relation to the power-transfer threshold
□ Research challenges: Usefully increase power density
(i.e., do more work) without degrading performance
Slide 13
Intermittency
1. Datacenter Model
2. Wind Intermittency
3. Renewables in the DC
4. Related Work
■ Goal: Understand impact of
□ 1. power-transfer threshold
□ 2. power consumption
■ Approach: Simulate the behavior of
the ATS under different
datacenter consumption
scenarios
Battery Backup
ATS
Grid Energy
STS/
UPS
ATS
Off-grid renewables
Primary
■ Challenge: Intermittency Renewable
Sources
□ We used traces from
National Renewable
Energy Lab
□ 10 min. snapshots of power
production at turbines in
MT and CA
□ Normalized to a 1MW turbine
Power Transfer System
1 MW
Site
MT
MT
1 MW peak
Secondary
Output
Time
Avg. Power(MW)
12:10am
3.0
12:20am
2.9
Slide 14
■ How precious are renewables?
■ Plot intermittent outages and
continuous power
■ Battery backup too costly
■ But enough continuous supply
for renewable-aware
management
□ Some power production 42%
of the time for MT turbine
□ Peak power 25% of the time
□ Peak renewables are 40%
percent more precious
Prob. for 10 min-width bins
1. Datacenter Model
2. Wind Intermittency
3. Renewables in the DC
4. Related Work
50%
MT TurbineNo Production (outages)
25%
Prob. for 10 min-width bins
Intermittency
8 MWh / 200 Wh/kg LI =
40000 kg LI or ~$3mil
[wikipedia, Lithium Ion]
0%
10
100
300
>500
Duration (minutes)
50%
MT TurbineCont. Power Production
25%
0%
10
100
300
>500
Duration (minutes)
Slide 15
■ NREL data provided input to our
datacenter model
■ Metric: Renewable energy used
□ MT produced 3.5M KW-hours
of renewables and CA
produced 2.3M KW-hours
(black line)
1. Datacenter Model
2. Wind Intermittency
3. Renewables in the DC
4. Related Work
4
KWh x 106
Intermittency
Powered by the MT wind turbine
3
2
1
0
100 KW
400 KW
700 KW
Average Power Consumption
Power-transfer threshold = zero
■ 1MW Power-transfer threshold
caused up to 65% drop
■ Zero power-transfer threshold
complicates the benefits of
increased consumption
KWh x 106
Power-transfer threshold = peak power
3
Powered by the CA wind turbine
2
1
0
100 KW
400 KW
700 KW
Slide 16
Intermittency
1. Datacenter Model
2. Wind Intermittency
3. Renewables in the DC
4. Related Work
■ Metric: Cost per KW-hour
■ Average price for commercial electricity $0.10 KW-hour
■ $2.4M to erect a wind turbine that is connected
(directly) to a datacenter [European Wind Energy
Assc.]
□ $1.6M installation
□ 2% annual maintenance fees
□ Lifetime of turbine: 20 years
■ Datacenter at CA or MT could use 24M KWh
□ Either high power consumption or zero threshold
□ Wind-powered datacenter in MT: $0.04 KWh
Slide 17
Intermittency
1. Datacenter Model
2. Wind Intermittency
3. Renewables in the DC
4. Related Work
■ Power-transfer system can improve server efficiency
□ PowerNap moves power supplies to the node level
[ASPLOS 2009]
• Improves energy proportionality
• Improves efficiency
□ Power Routing adds redundancy across PDUs
[ASPLOS 2010]
• Improves energy proportionality
• Workload spikes are co-located
■ Our work considers sustainability
□ Energy transfer versus power transfer
□ This work doesn't distinguish energy sources
Slide 18
Outline
□ Studying the Impact of Intermittency
□ Managing Renewables at the Request Level
• Request-Level event profiling [shen-asplos-2008]
• Event-driven power modeling [bellosa-ew-2000]
• Energy and Power Characteristics of Requests
• Load balancing scenario
□ Conclusion
Slide 19
Managing Renewables
1. Request Characteristics
2. Scenario
3. Future Work
Prev. section: MT wind-powered Datacenter with high
average power consumption (when renewables are
available) can potentially reduce costs to $0.04 Kwh
■ Idea: Use request-level management to increase power
consumption when only renewables are avaialable
• Route more requests to datacenters with unused
renewables
• Delay long-running Map-Reduce requests
□ Need mechanism to predict power/energy
consumption of a request
Slide 20
Managing Renewables
1. Request Characteristics
2. Scenario
3. Future Work
■ Request workloads executed in isolation
■ WattUp power meter measures watts and joules
■ Processor was not adjusted during tests
□ Disk accounted for as static consumption
Another case for energy proportional servers
Slide 21
Managing Renewables
■ Consider 2 futuristic datacenters
1. Request Characteristics
2. Scenario
3. Future Work
Power Transfer System
ATS
STS/
UPS
□ Server power is energy proportional
Load
Balancer
□ 1 datacenter uses renewable energy
■
Power monitoring system notifies when
precious renewables are available
50% peak
power util.
ATS
ATS
■
■
50% peak
power util.
Power Transfer System
ATS
STS/
UPS
Challenges
Load
Balancer
□ DCs connected by hard-to-change
DNS
□ Need optimization procedures
[K. Le-HotPower-2009]
0.5
STS/
UPS
Load balancer increases the renewablepowered datacenter's workload
□ Uses request power models as
guidance
0.5
40% peak
power util.
0.2
ATS
ATS
0.8
STS/
UPS
70% peak
power util.
Slide 22
Managing Renewables
1. Request Characteristics
2. Scenario
3. Future Work
Our ongoing research is on the design, management,
and mechanisms that will increase the use of
precious renewables in the datacenter
1. Study trade offs in the configuration of the power-transfer
system, i.e. ATSs, PDUs, and UPSs
□ Amount of precious renewables transferred to compute equip.
□ Brownouts
□ Practicality—space, cost, and increment-ability
2. Study load balancing policies with renewable-aware
mechanisms
e.g., per-request vs. per-session round robin, feedback based
□ Redirection speed for intra- and inter-datacenter workload
Slide 23