Introduction
Data centers are significant energy consumers, accounting for approximately 2% of total electricity usage in the United States. To mitigate this environmental impact, researchers develop advanced algorithms for resource allocation, scheduling, and pricing. However, the efficacy of these algorithms depends heavily on the accuracy of the underlying demand models.
Historically, workload models have treated resource attributes like CPU and memory usage in isolation. In reality, these attributes are often correlated; for instance, a job with high CPU demands frequently has specific memory requirements. Ignoring this joint distribution leads to inefficient scheduling and wasted energy. Real-world cloud data is notoriously "messy"—heavy-tailed and non-linear—making it difficult to fit with standard statistical distributions.
Methodology: The Copula Approach
The authors propose a two-step framework using Copulae to model the joint distribution of CPU and memory usage:
- Marginal Modeling: First, the best univariate distribution for each attribute is identified. The Burr XII distribution (also known as the Singh-Maddala distribution) was selected for its flexibility in modeling both decreasing and bell-shaped curves with heavy tails.
- Copula Integration: A copula function is then used to describe the dependence between these variables independently of their marginal distributions. The study evaluated several types, including Gaussian and Archimedean copulae (Gumbel, Frank, and Clayton).
- Validation: The model's accuracy was verified using a Parametric Bootstrapped Kolmogorov-Smirnov (KS) test, a rigorous method to determine if the observed data aligns with the hypothesized model.
Data and Marginal Analysis
The study utilized the Google cluster-usage traces (2011), encompassing 29 days of workload data from a production cluster of approximately 11,000 machines. The analysis focused on approximately 760,000 successfully completed processes, specifically targeting the "bulk" of the workload (the bottom 82% of usage) to avoid extreme outliers that would necessitate more complex multimodal models.
A key finding was the superior performance of the Burr XII distribution. While traditional models like Log-Normal or Pareto failed goodness-of-fit tests (p-values of 0), Burr XII accurately captured the data's skewness and kurtosis, providing a high-fidelity fit for the marginal distributions.
Key Results
- High-Fidelity Marginals: The Burr XII distribution achieved high p-values: 0.52 for CPU and 0.74 for memory. In statistical testing, a high p-value indicates that the model is indistinguishable from the real data.
- Optimal Joint Fit: The Frank copula was identified as the most effective function for coupling the variables. The complete model (Burr XII marginals + Frank copula) achieved a p-value of 0.44, representing a high-fidelity joint demand model.
- Correlation Significance: The research demonstrated that CPU and memory usage are significantly correlated, particularly for very small or very large jobs. Modeling them as independent variables results in a significant loss of model fidelity.
Conclusion and Future Directions
The copula-based framework offers a modular approach to workload modeling. Researchers can adapt the model to different data center environments by swapping marginal distributions or coupling functions. While this study focused on the bivariate case, the machinery generalizes to higher dimensions (e.g., including I/O or network bandwidth).
Future work aims to address current limitations, such as:
- Outlier Modeling: Implementing "mixture models" to capture rare, resource-heavy extreme events currently excluded from the "bulk" analysis.
- Dimensionality: Extending the model to include additional attributes, which becomes computationally expensive in high-dimensional spaces.
- Temporal Dynamics: Integrating these spatial joint distributions with time-series analysis to model how demand fluctuates over time.
By providing a more accurate representation of multi-attribute demand, this research enables the development of more energy-efficient cloud infrastructure.
Summary updated