parrobaartecsecs_11-03-13
Post on 25-Dec-2014
52 Views
Preview:
DESCRIPTION
TRANSCRIPT
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
DVFS STUDY IN DATA CENTER INFRASTRUCTURES
ENERGY EFFICIENCY APPLICATION
Patricia Arroba García
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Contents
1. Research objective
2. Context
3. Motivation
4. Energy model
5. Energy profiling
6. Results
7. Conclusions
8. Future work
2 Patricia Arroba | DVFS study in Data Center infrastructures
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Research objective
• Reducing energy consumption in Cloud computing data centers
– Application awareness
– Thermal behavior awareness
• Management of low power modes (DVFS)
• Resource selection and consolidation
• Task scheduling
3 Patricia Arroba | DVFS study in Data Center infrastructures
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Research objective
• DVFS in general purpose computing represents savings about 70%
• Energy consumption in Data Centers is dominated by consumption of idle hosts
• In practice, Data Centers do not use DVFS
4 Patricia Arroba | DVFS study in Data Center infrastructures
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Context
• State of the art: DVFS policies *
– frequency Dynamic energy consumption
• Consequence
task execution Static Energy consumption
5
Patricia Arroba | DVFS study in Data Center infrastructures
*E. Quintana: Considers this consequence in his research “Energy-Aware Matrix Computations on Multi-Core and Many-core Platforms”
time
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Context
• State of the art: Energy consumption models
– Dynamic energy = f ( f, V, CPU load %)
– Static energy = f ( T or T3) fitting complexity
• Consequence
– Weak static energy consumption model
frequency and voltage dependencies missing
simplified temperature dependance
6
Patricia Arroba | DVFS study in Data Center infrastructures
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Motivation
Practical evaluation in real servers
• Validation of a complete energy model
• Analyzing the impact of:
– CPU load %
– Temperature
– DVFS modes & execution time
on the overall energy consumption
7
Patricia Arroba | DVFS study in Data Center infrastructures
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Energy model
8
(Taylor Series)
Patricia Arroba | DVFS study in Data Center infrastructures
m: host i: task k: DVFS mode
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Energy model
9
Leakage energy of the rest of the server (NOT CPU) It does not depend on the DVFS mode. Fixed frequency.
Patricia Arroba | DVFS study in Data Center infrastructures
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Energy model
10 Patricia Arroba | DVFS study in Data Center infrastructures
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Energy Profiling
• Server: Intel Xeon RX300 – 1 CPU, 8 threads hardware
– 7 DVFS modes 1.73 – 2.4 GHz
– Tools: Ipmitool (PTOT, TCPU), perf (IPC), ps(CPU%)
cpufrequtils (DVFS modes)
• Workload: Lookbusy – Synthetic workload
– Highly controllable (CPU load %)
– Does not affect other devices (memory, disk, ... )
11 Patricia Arroba | DVFS study in Data Center infrastructures
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Curve fitting
• Matlab fitting functions
– Maximum fitting error = 13.61 %
– Mean fitting error = 2.02 %
12 Patricia Arroba | DVFS study in Data Center infrastructures
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Model application results:
Dynamic Energy vs load CPU%
13 Patricia Arroba | DVFS study in Data Center infrastructures
Wh
%
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Model application results:
Leakage vs load CPU
14 Patricia Arroba | DVFS study in Data Center infrastructures
Wh
Hz
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
15
Model application results:
Leakage vs temperature
Patricia Arroba | DVFS study in Data Center infrastructures
Wh
Hz
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
• Temp: 303 K
16
Model application results:
Energy vs DVFS (CPU 10%)
• Temp: 315 K
Difference (Temp.) 5%
Difference (DVFS) 12% Difference (DVFS) 13%
Patricia Arroba | DVFS study in Data Center infrastructures
Wh Wh
Hz Hz
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
17
Model application results:
Energy vs DVFS (CPU 100%)
• Temp: 303 K • Temp: 315 K
Difference (Temp.) 5%
Difference (DVFS) 5.5% Difference (DVFS) 6.8%
Patricia Arroba | DVFS study in Data Center infrastructures
Wh Wh
Hz Hz
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Conclusions
18 Patricia Arroba | DVFS study in Data Center infrastructures
Energy-Aware Matrix Computations on Multi-Core and Many-core Platforms. E. Quintana
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Conclusions
• Energy model validated in a real server
• The state of the art does not take into account
the execution time penalty in the static energy
consumption
– Consumption only activating CPU: Best savings for the highest operating frequency
• Promising future work
19 Patricia Arroba | DVFS study in Data Center infrastructures
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Future work
• Static energy consumption masks CPU dynamic consumption so the static contribution dominates the DVFS policy
• Memory consumption is significant as its contribution will increase the percentage of dynamic consumption
20 Patricia Arroba | DVFS study in Data Center infrastructures
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Future work
• Due to this reason, an optimum DVFS mode could be found to increase energy savings
21
Fixed Optimum
Patricia Arroba | DVFS study in Data Center infrastructures
Application features
CPU % Memory demand %
DVFS optimum
Wh
Hz
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Future work
• Real applications can be characterized by its CPU and memory demand
• DVFS modes can be adjusted to these features to increase the overall data center energy efficiency considering the application
• We have developed a MILP to simulate the allocation of tasks, the resource selection and configuration in a data center infrastructure that will be revised to reflect this new target
22 Patricia Arroba | DVFS study in Data Center infrastructures
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Publications
• Breaking the Energy-Wall in e-Health Scenarios with a Novel Computing Paradigm. M. Zapater, P. Arroba, J.L. Ayala, J.M. Moya, K. Olcoz, C.A. López-Barrio, FGCS 2013.
23 Patricia Arroba | DVFS study in Data Center infrastructures
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
Thanks for your attention
24 Patricia Arroba | DVFS study in Data Center infrastructures
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
25 Patricia Arroba | DVFS study in Data Center infrastructures
“Ingeniamos el futuro”
CAMPUS OF
INTERNATIONAL
EXCELLENCE
26 Patricia Arroba | DVFS study in Data Center infrastructures
top related