Energy Demand Response Modeling for High Performance Computing - - PowerPoint PPT Presentation

Energy Demand Response Modeling for High Performance Computing Systems

Kishwar Ahmed and Jason Liu Florida International University

Workshop on Modeling and Simulation of Systems and Applications August 15-17, 2018u u University of Washington, Seattle, Washington

Demand Response

• Participants reduce energy

consumption during

• Emergency events
• High electricity price period
• Emergency demand response
• Mandatory energy reduction to

target level

• Economic demand response
• Voluntary participation based on

economic incentives

DR signal DR signal Energy reduction

2

Why Demand Response?

Ø Increase in demand response participation

v Many well-known companies, such as Google, Apple, etc. v Participation in demand response to double in 2020

Financial benefits Environmental benefits Power system stability

3

HPC is Energy-Costly!

• Worldwide investment on

supercomputers

• In 2016: $38 billion
• Supercomputer’s lifelong

energy cost almost equals investment cost

• Advent of Exascale
• 20MW à $20 million/year

for electricity

Hardware 36% Power 28% Staff 14%

Mainteinance 14% Other 8%

4

Source: “Total Cost of Ownership in High Performance Computing. HPC data center cost considerations: investment, operation and maintenance.” in SoSE 2014

HPC in Demand Response

• Can HPC systems reduce the energy consumption and

energy cost through emergency and economic demand response participation?

• Supercomputers are willing to participate [Patki et al., 2016;

Bates et al. 2015]

• Our solutions:
• Emergency demand-response model
• Application performance loss vs. energy reduction gain?
• Economic demand-response model
• How to incentivize HPC users for demand response

participation?

5

Emergency DR Model

• Power/performance prediction model
• Empirical data
• Polynomial regression
• Demand response job scheduling
• FCFS with possible job eviction (to ensure power limit)
• Resource provisioning
• DVFS, power capping, node scaling

6

Power/Performance Prediction Model

Apply regression (quite a few alternatives) on power and execution time

50 100 150 200 250 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 Average Power (Watt) CPU Frequency (GHz) Quantum ESPRESSO Gadget Seissol WaLBerla PMATMUL STREAM 20 40 60 80 100 120 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 Execution Time (Min) CPU Frequency (GHz) Quantum ESPRESSO Gadget Seissol WaLBerla PMATMUL STREAM 100 200 300 400 500 600 700 800 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 Energy Consumption (KJ) CPU Frequency (GHz) Quantum ESPRESSO Gadget Seissol WaLBerla PMATMUL STREAM

7

Job Scheduling

• During normal operation:
• Traditional job scheduling
• Optimized for best performance (max frequency)
• During demand response period:
• Minimize energy for resource allocation
• DVFS, power-capping, node scaling
• Reduce power limit
• May have to evict some jobs

8

Resource Allocation

• During normal operation
• Run applications at maximum frequency for best

performance

• During demand response: energy conservation

9

Model Evaluation

Vary system size: 128, 256, and 512 processors

200 220 240 260 280 300 128 256 512 Average Energy (KJ) Number of Processors Performance-policy Demand-response (DR Event) Demand-response (Non-DR Event) Powersave-policy 1000 1500 2500 3500 4500 5500 128 256 512 Average Turnaround Time (s) Number of Processors Performance-policy Demand-response (DR Event) Demand-response (Non-DR Event) Powersave-policy

Reduced energy consumption at moderate increase in turnaround time

10

Economic DR Model

• Economic demand response
• Voluntary participation based on economic incentives
• How to incentivize HPC users for participation?
• Participation may introduce execution delays
• Need a proper rewarding mechanism
• HPC economic DR model
• A contract-based rewarding mechanism to incentivize

HPC users’ participations

11

Contract Theory

• A formal (economic) study to develop contracts

between parties

• Principal: who offers the contracts (HPC operator)
• Agents: who are offered the contracts and can accept/

reject (HPC users)

• Widely used in theory and practice
• Economics (e.g., managerial compensation)
• Communication (e.g., cellular network)

12

An High-Level Example

($1220, 15) Type#1 Type#2 Type#3 Job types User#1à Type#2 User#2à Type#1 User#3à Type#3 HPC Users User’s utility maximized when selects own type’s contract HPC operator

13

Resource Allocation

14

Energy and Reward

200 400 600 800 1000 1200 1400 1600

12345678910 11 12 13 14 15 16 17 18 19 20 21 22 23 24

Energy Reduction (KWh) Time (Hour) Type#1 Type#2 Type#3 Type#4 Type#5 Type#6 2 4 6 8 10 12 14 16 18

1 2 3 4 5 6 7 8 9 1 1 1 1 2 1 3 1 4 1 5 1 6 1 7 1 8 1 9 2 2 1 2 2 2 3 2 4

Reward ($) Time (Hour) Type#1 Type#2 Type#3 Type#4 Type#5 Type#6

Energy reduction and rewards throughout entire time periods

15

Conclusions

• HPC demand-response models
• Emergency demand response participation
• Economic demand response participation
• A win-win situation to all:
• HPC systems reduce energy cost
• HPC users earn rewards
• Power grid achieves energy reduction and power

stability

16

Thanks!

Kishwar Ahmed, Jason Liu, ModSim Workshop, August 2018

Acknowledgements:

17