workload-based prediction of cpu temperature and usage for small-scale distributed systems
TRANSCRIPT
Workload-Based Prediction of CPU
Temperature and
Usage for Small-Scale Distributed Systems
1, 2, 4, 5 Bangladesh University of Engineering Technology3 University of Nebraska, Lincoln, USA
The 4th International Conference on Computer Science and Network Technology (ICCSNT), Heilongjiang University,
HARBIN, CHINA
Raju Ahmed Shetu1, Tarik Toha2, Mohammad Mosiur RahmanLunar3, Novia Nurain4, A. B. M. Alim Al Islam5
Outline
• Background and motivation
• Related work
• Proposed models
▫ Empirical analysis
▫ Mathematical formulation
▫ Validation
• Future work and conclusion
22
Background
Source: Moss, David L. "Data center operating temperature: the sweet spot." A Dell Technical White Paper. Round Rock, Texas (2011). 3
27
28
29
30
31
32
33
34
62 66 70 74 77 80 84 88 91 95
Po
wer
(k
W)
Temperature in F
Composite server power
Composite server
power (i.e. energy
and airflow
consumption)
increases !!!
0
0.5
1
1.5
2
2.5
15 17.5 20 22.5 25 27.5 30 32.5 35 37.5 40 42.5 45
Fa
ilu
re r
ate
Temperature in C
Background(contd.)Server failure rate
4
Server failure increases with an
increase in temperature !!!
Source: Kelley, Chris, Cisco Harkeeret Singh, and Vic Smith. "Data center efficiency and IT equipment reliability at wider operating temperature and humidity"
MotivationEffective thermal
management is
required for the
efficacy of a
system
Proper usage of
resources is
necessary to
ensure reliable
QoS
Motivation
We need to develop mathematical models for forecasting
effectively the future CPU temperature and usage to derive
temperature-efficient resource allocation policies and
workload scheduling algorithms
• Coskun et al. [IEEE Transactions on CADICS, 2009]-Exploit auto-regressive moving average (ARMA) for thermal model
• Yeo et al. [Design Automation Conference, 2008 ]- Exploit generic polynomial function
• Ramos et al. [HPCA, 2008]- Exploit thermodynamic laws- Based on macro-level parameters such as (core temperature/utilization, airflow
velocity, dynamic voltage and frequency scaling per core and processor
• Moore et al. [ICAC, 2006]- Utilizes a neural network to learn and predict temperature under static workload
• Ayoub et al. [IEEE Transactions on CADICS, 2011]- Workload-independent temperature predictor based on the band limited
property of the temperature
7
Related Work
• Li et al. [KDD, 2011]
- Exploit CFD modeling to predict CPU temperature based on principles from thermodynamics and fluid mechanics
• Li et al. [VLDB Endowment, 2012 ]- Deploy machine learning techniques to perform
resource (CPU usage) estimation and prediction
8
Related Work
Our Contribution in this Paper
We develop simple prediction models of CPU temperature and usage considering impacts of
macro-level parameters such as dynamic workload and different number of machines
9
Experimental Analysis
Experimental setup
10
Parameter Values
CPU Intel Core 2™Duo
Clock speed 2.40GHz
Memory 2 GB
Workload 2.9GBto 58.7GB
# machines 6 to 16
Experimental Results
Variation in CPU temperature in response to a change in the
number of machines
11
Exhibits minimal response to a change in
the number of machines
Experimental Results
Variation in CPU temperature in response to a change in the
size of workloads
12
Exhibits significant response to a change in
workloads
Validate the prediction of CPU temperature basedon only the workloads
Experimental Results
Variation in CPU usage in response to a change in the size of
workloads
13
Exhibits significant response to a change in
workloads
Validate the prediction of CPU usage basedon only the workloads
Proposed Models
We formulate the equation from the data of experimental results using Matlab
Workload-based prediction model for CPU temperature can be represented as followed:
t = P1 x2 + P2 x + P3
Here,
t=CPU temperature in °C
x= the size of workload in GB.
14
Proposed Models
Workload-based prediction model for CPU usage can be represented as followed:
u = a0 +
Here,
u= CPU usage in %
x= the size of workload in GB.
15
)sin()( cos4
1
xibxia i
i
i
Validation of Proposed Models
16
Our model exhibits
close match
Validation of Proposed Models
17
Our model mostly
captures the trend
Future Work
We plan to make our proposed models more rigorous through investigating impacts of other parameters
such as external temperature, number of cooling fans, and dynamic voltage and
frequency scaling per core and processor, etc.
18
Conclusion• Existing prediction models fail to adopt dynamic
changes in workload
• We focused on developing workload-based prediction models for CPU temperature and usage
• We confirm accuracy of our proposed models by comparing the values estimated from our models against the measurements obtained from a real implementation
19
Any QuestionsEmail: [email protected]
20
References
1. Ginter, Karl L., Victor H. Shear, Francis J. Spahn, David M. Van Wie, and Robert P. Weber.“Trusted infrastructure support systems, methods and techniques for secure electroniccommerce, electronic transactions, commerce process control and automation, distributedcomputing, and rights management.” U.S. Patent 8,185,473, issued May 22, 2012.
2. Stewart, Brett B., James Thompson, and Kathleen E. McClelland. “Distributed networkcommunication system to provide wireless access to a computing device at a reduced rate.”U.S. Patent 8,588,130, issued November 19, 2013.
3. Qu, J., Yang, J., Liu, H., and Wu, R. “Clock Synchronization in Air Traffic Control SystemUsing NTP”. International Journal of Digital Content Technology and its Applications, vol.7(1), pp. 344, 2013.
4. Kopetz, H. “Real-time systems: design principles for distributed embedded applications”.Springer Science and Business Media, 2011.
5. Coskun, A., Rosing, T., and Gross, K. “Utilizing Predictors for Efficient Thermal Managementin Multiprocessor SoCs.” IEEE Transactions on Computer-Aided Design of IntegratedCircuits and Systems, vol. 28(10), pp. 1503–1516, 2009.
6. Coskun, A., Rosing, T., and Gross, K. “Proactive Temperature Management in MPSoCs.” InInternational Symposium on Low Power Electronics and Design, pp. 165–170, 2008.
7. Yeo, I., Liu, C., and Kim, E. “Predictive Dynamic Thermal Management for Multi-coreSystems.” In Design Automation Conference, pp. 734–739, 2008.
21
References
8. Ramos, L., and Bianchini. R. “C Oracle: Predictive thermal management for data centers.” In IEEE 17th International Symposium on High Performance Computer Architecture (HPCA), pp. 111–122, 2008.
9. Moore, J., Chasey, J., and Ranganathanz, P. “ Weatherman: Automated, online, and predictive thermal mapping and management.” In Proceedings of the 3rd IEEE ICAC, pp. 155–164, 2006.
10. Choi, J., Kim An, Y., Srebric, J., Wang, Q., and Lee. J. “Modeling and managing thermal profiles of rack mounted servers with thermostat.” In Proceedings of the 13th International Symposium on High-Performance Computer Architecture, pp. 205–215, 2007.
11. Li, L., Liang, C. J. M., Liu, J., Nath, S., Terzis, A., and Faloutsos, C. “Thermocast: a cyber physical forecasting model for data centers.” In ACM KDD, 2011.
12. Ayoub, R., Indukuri, K., and Rosing, T. S. “Temperature Aware Dynamic Workload Scheduling in Multisocket CPU Servers.” IEEE Transactions on computer aided design if integrated circuits and systems, vol. 30(9), pp. 1359, 2011.
13. Li, Jiexing, et al. “Robust estimation of resource consumption for SQL queries using statistical techniques.” Proceedings of the VLDB Endowment vol. 5(11), pp. 1555–1566, 2012
22