Slides - Computer Science

Report
Parasol and GreenSwitch:
Managing Datacenters Powered by
Renewable Energy
Íñigo Goiri, William Katsak, Kien Le,
Thu D. Nguyen, and Ricardo Bianchini
Department of Computer Science
Motivation
• Datacenters consume large amounts of energy
• High energy cost and carbon footprint
– Brown electricity: coal and natural gas
• Connect datacenters to green sources: solar, wind
Apple DC in Maiden, NC
40MW solar farm
2
Challenges and opportunities
Variable
Solar power
Storage?
Load
Power
Workload
Time
Source?
• Scheduling workload/energy sources
– Lower costs: brown energy, peak brown power, capital
• Study opportunities in green datacenters
– Build hardware/software
3
Outline
• Motivation
• Parasol
– Solar-powered micro-datacenter
• GreenSwitch
– Manage workload and energy sources
• GreenSwitch results
• Conclusion
4
Parasol
16 solar panels: 3.2kW
2 inverters: DC→AC
16 lead-acid batteries: 32kWh
2 charge controllers: Off-grid, DC↔AC
Installed on the roof
• Steel structure
• Small container
IT equipment
• 2 racks
• 64 Atom servers
• 2 switches
• 3 PDUs
5
Electrical infrastructure
DC
Inverter
AC
Electrical
Panel
AC
AC
AC
Power grid
Charge
Controller
DC
IT Equipment
6
Power
Example energy source management
DC
Inverter
Time
AC
Electrical
Panel
AC
AC
AC
Power grid
Charge
Controller
DC
IT Equipment
7
Power
Example energy source management
DC
Inverter
Time
AC
Electrical
Panel
AC
AC
AC
Power grid
Charge
Controller
DC
IT Equipment
8
Power
Example energy source management
DC
Inverter
Time
AC
Electrical
Panel
AC
AC
AC
Power grid
Charge
Controller
DC
IT Equipment
9
Possible energy source management
Power
Basic
Time
10
GreenSwitch
• Minimize brown electricity cost
– Brown energy
– Peak brown power
– Battery lifetime constraint
GreenSwitch
• Manage energy sources
– Use solar/net metering
– Charge/discharge battery
– Limit brown peak power
• Manage workload
Get status
Perform changes
Parasol
– Turn servers on/off
– Delay deferrable jobs
11
GreenSwitch architecture
Battery
Charge Level
Energy Source
Schedule
Solver
Configurer
Workload
Prediction
Energy
Availability
Prediction
Solar
Brown
Price
Workload
Schedule
Model &
MILP
Workload
Time
Power
Power
Predictor
GreenSwitch
Workload
Time
Parasol
Use solar to power all servers
Charge battery with surplus solar
12
Experimental environment
• Evaluation on 64 Parasol nodes
– 12 one-day experiments
– Deferrable vs. non-deferrable workloads
– Baseline datacenter (no solar, no batteries, no delays)
• New Jersey brown electricity pricing
– On/off-peak energy, peak power, net metering
• GreenSwitch for Hadoop (configurer)
13
Experimental environment
IT load
SWIM: Facebook based workload [MASCOTS’11]
14
Parasol without GreenSwitch
Green available
IT load
Net metering
Green use
Brown use
66% cost savings → Solar amortized in 7 years
15
GreenSwitch: non-deferrable workload
Green available
IT load
Battery charge
Battery discharge
Peak grid power
Green use
Brown use
75% cost savings → Batteries cannot be amortized
16
GreenSwitch: deferrable workload
Green available
Net metering
Battery charge
IT load
Battery discharge
Green use
96% cost savings → Solar + batteries amortized in 7.6 years
17
Parasol: a real system
•
•
•
•
Real software running on real hardware
Power losses
Overhead of energy source switching
System limitations
– Net metering vs. Battery charging
– Use brown vs. Net metering
– Green battery charging vs. Use brown
18
Conclusions
• Green datacenters
– Challenges & opportunities
– Hardware/software solution
• GreenSwitch benefits
– Delaying load and solar gives the best results
– Reduces amortization time by 1.8-2x
– Flexibility: no batteries, workloads, wind…
19
Parasol and GreenSwitch:
Managing Datacenters Powered by
Renewable Energy
Íñigo Goiri, William Katsak, Kien Le,
Thu D. Nguyen, and Ricardo Bianchini
http://parasol.cs.rutgers.edu
• Fewer conversions
– Grid-tie: AC→DC
• Low conversion losses
• Commodity hardware
– Availability
– Cheaper
• Cheaper transmission
21
Parasol: lessons learned
• Not cheap
– Flexibility adds complexity and cost
– Placement on the roof
– Full monitoring
• Not easy
– Complete design
– Dealing with facilities, companies, and vendors
– Mistakes are easy
– Delays, delays, and more delays
22
Viability of green datacenters
• Example companies investing in self-generation
– Apple, Microsoft, McGraw-Hill
• Research shows cost savings
• Getting more attractive
– Space: Solar panels are getting more efficient
– Cost: 7x cheaper than 30 years ago
– Governmental Incentives
• Grid-centric: power purchase agreement
– Losses, dependence, long term scalability
23
Other results
• Energy storage
– Batteries
– Net metering pricing
• Peak grid power charges
• Solar availability
– Cloudy, sunny, rainy…
• Other workloads
– Nutch
• Grid outage
24

similar documents