A methodology for full system power modeling in heterogeneous data centers
- 1. www.bsc.es UCC 2016 Shanghai, 6th November Mauro Canuto, Raimon Bosch, Mario Macías, Jordi Guitart A Methodology for Full-System Power Modeling in Heterogeneous Data Centers
- 3. 3 Objectives • A new methodology to model power consumption that is platform and application agnostic. • Overcome existing limitations: platform and application dependencies, inaccuracy, complexity, … • Approximate energy with resource usage indicators and its correlations. • Present our final results.
- 4. 4 Introduction • Why do we need power modellers? – Energy awareness. – Increasing use of energy in datacenters. – Not access to power consumption meters. • What are the limitations of current energy modellers? – Considering only the CPU. – Inadequate selection of indicators. Correlations and nonlinear behaviors. – Use a training that relies on architecture-specific indicators.
- 5. 5 Scope • How is our methodology? (1 / 3) - Derives full-system power models by considering the impact on energy consumption of all the resources in the system, namely processor, cache, memory, disk, and network, and can be easily extended to include other resources (i.e. GPUs) as needed. - Derives a single model per platform that works with high accuracy for heterogeneous applications with different patterns of resource usage and energy consumption. - Derives models by using a minimum set of resource usage indicators, which are selected for each platform according to their correlation with the power consumption.
- 6. 6 Scope • How is our methodology? (2 / 3) - Derives models that capture non-linear relations between the resource usage and the power consumption. - Exploits machine learning techniques to drive the selection of resource usage indicators and the capture of non-linear relations without human intervention. - Gets full-system information needed to generate the models by means of a comprehensive monitoring framework that seamlessly integrates several data sources, including system information, performance counters, and overall power consumption measures.
- 7. 7 Scope • How is our methodology? (3 / 3) - Uses the same training set for each platform, which is independent of the applications used to validate the models and comprises various micro-benchmarks that selectively stress each resource at varying levels of utilization. - Validates the models with real applications, which are commonly found in Cloud data centers, with different patterns of resource usage and energy consumption. - Achieves platform- and application-agnosticism by using the same tools and systematically following the same steps to derive power models in heterogeneous platforms.
- 8. COLLECTION, TRAINING, MODEL GENERATION & VALIDATION
- 10. 10 Training micro-benchmarks • Stress different components of the target platform (CPU, cache, main memory, network, and disk) at different intensity levels • CPU: Stress-ng, Ibench, Sysbench, Prime95 and Linpack-neon. Increase intensity until you hit 100% of resources. • Memory: Stress-ng, Pmbw, STREAM. • Disk: Stress-ng, Fio • Network: iperf increasing bandwidth by steps of 10%.
- 12. 12 Model generation - An initial selection of features is obtained by excluding those not showing significant variance in the training experiments. - For each resource usage indicator on each micro-benchmark, we perform a regression analysis to find the best the relation METRIC vs. CONSUMPTION. - Apply logarithmic and polynomial transformations: ● If a logarithmic function returns a best correlation than the original function, we apply the log(metric) transformation to the data corresponding to that metric. ● If the relation fits better with a polynomial function, we apply the pow(metric, α) transformation to the data. The α parameter is obtained through the nls function from stats package for R language, which determines non-linear least-squares estimates of non-linear models.
- 13. 13 Model generation l CPU cycles for the Prime95 micro-benchmark in srv-xeon-2 platform is best approximated by using a logarithmic function. => l training dataset = l Filter the selected features. l Join datasets of all microbenchmarks. l Perform the polynomial and logarithmic transformations. l Use linear regression in order to fit our model.
- 14. 14 Model validation l Cloudsuite: Benchmarks based on real cloud applications + HPC validation with NAS Parallel Benchmarks.
- 15. RESULTS
- 16. 16 Results
- 17. 17 Results: Intel Xeon ● Validate against Lasso regression to automatically incorporate features ○ linear models (LNLS) ○ nonlinear polynomial models (PLLS) ○ Our model (OURS) ● The average MAPE of our models is 4.3% for srv-xeon-1 and 5.7% for srv-xeon-2. ● Very accurate considering the range of applications.
- 19. 19 Results: AMD Opteron ● High range of power consumption that can be generated by the two platforms (up to 400 W for srv-opt-1 and 360 W for srv-opt2) ● The average error of our models is still very low: 5.4% for srv-opt-1 and 5.7% for srv-opt-2.
- 21. 21 Results: Low power ● Intel Atom (srv-atom) & ARM Cortex-A (srv-arm). ● Very promising (as shown in Table 4). The average MAPE for srv-atom is 2.6% and for srv-arm is 5.2%. ● Lasso models provide good accuracy for NAS HPC jobs, but they show important errors for CloudSuite benchmarks.
- 23. 23 Conclusions - A platform and application-agnostic methodology for full-system power modeling in heterogenous data centers. - Demonstrated with different power consumption profiles ranging from high- performance to low-power architectures. - Models provide high accuracy (around 5% of average estimation error).
- 24. www.bsc.es Thank you! For further information please contact raimon.bosch@bsc.es 24