Title: EnergyAware Duplication Strategies
1 Brian Stinar
IEEE Cluster 2006
- Energy-Aware Duplication Strategies
- for Scheduling Precedence
- Constrained Parallel Tasks on Clusters
2Overview
- Motivation of our research
- System Models
- Algorithms EADUS TEBUS
- Simulation Results
3Motivation
- Rapidly increased applications
- (commercial and scientific)
- Over consumed energy in cluster systems
- (a large-scale cluster commonly requires 40TWh
per year, costing over 4B per year at the price
of 100 per MWh)
4 System Models
Homogeneous
5 System Models
A parallel application with a set of precedence
constrained tasks is represented in form of a
Directed Acyclic Graph (DAG), modeled as a pair
(V, E).
6 System Models
7 Scheduling Algorithms
Linear Schedule
Time 39s Energy 234J
Time 32s Energy 242J
No Duplicate Schedule
Time 29s Energy 284J
Task Duplicate Schedule
CPU_Energy_Busy6W Link_Energy_Busy1W
8EADUS TEBUS Algorithms
9If duplicate T1 More_energy 48-6 42J
Save_time 6S Ratio 42/6 7 If we set
threshold 10 EADUS will not duplicate T1
TEBUS will duplicate T1
CPU_Energy_Busy6W Link_Energy_Busy1W Here we
Ignore idle energy of processors and links
Time 32s Energy 242J
EADUS
10Simulation Results
Characteristics of System Parameters
11Simulation Results
(a) Communication Energy Consumption
(b) Total Energy Consumption
Fig. 1. CCR Sensitivity for Gaussian Elimination
12Simulation Results
(a) CPU Energy Cost in Different Thresholds
(b) Total Energy Cost in Different Thresholds
Fig. 2. Threshold Sensitivity for Gaussian
Elimination, CCR 0.7
13Simulation Results
(a) CPU Energy Consumption (b)
Total Energy Consumption
Fig. 3. CCR Sensitivity for Fast Fourier Transform
14Simulation Results
(a) Average Energy Cost of Gaussian Elimination
(b) Average Energy Cost of Fast Fourier
Transform
Fig. 4. Energy Evaluation for Gaussian
Elimination and Fast Fourier Transform
15Thanks to the rest of my research group
Dr. Xiao Qin Assistant Professor NMTZiliang
Borris Zong Ph.D. student NMTAdam Manzanares
MS student UC Boulder
16 Questions?