Technical Summary: SOFA – Sustainable Container Orchestration in Geo-Distributed FaaS Datacenters

1. Problem Statement

The paper addresses the challenge of scheduling Function-as-a-Service (FaaS) workloads across Geo-Distributed Datacenters (GDDCs). While GDDCs offer opportunities to reduce costs and environmental footprints by exploiting heterogeneity in renewable energy availability, electricity prices, and carbon intensity, they introduce significant complexity.

The core problem is the multi-objective optimization of FaaS scheduling under strict constraints. Existing approaches often focus on single objectives (e.g., minimizing energy consumption) or fail to holistically account for all environmental factors. Specifically, the paper identifies gaps in current research:

• SLA Heterogeneity: FaaS workloads have diverse and strict Service Level Agreements (SLAs), particularly regarding latency. Migrating workloads across geographic distances can increase end-to-end latency, leading to SLA violations.
• Incomplete Sustainability Metrics: Prior work often neglects water usage or fails to distinguish between energy consumption and carbon emissions (since low-carbon energy sources may still be expensive or water-intensive).
• Lack of Trade-off Flexibility: Many existing schedulers produce a single solution rather than a diverse set of Pareto-optimal solutions, limiting the ability of cloud providers to choose trade-offs between utility, cost, carbon, and water.

The goal is to co-optimize four conflicting objectives simultaneously:

1. SLA Violations (Utility Loss): Minimizing the loss of utility caused by missing latency deadlines.
2. Energy Costs: Minimizing operational costs based on Time-of-Use (TOU) pricing.
3. Operational Carbon Emissions: Minimizing emissions from both electricity generation and water usage.
4. Water Usage: Minimizing water consumed for cooling and electricity generation.

2. Methodology: The SOFA Framework

The authors propose SOFA (Sustainable Container Orchestration in Geo-Distributed FaaS Datacenters), a novel resource orchestration framework. SOFA integrates a metaheuristic approach with machine learning to generate a diverse Pareto front of solutions.

2.1 System Modeling

The framework relies on comprehensive models for the GDDC environment:

• Workload Model: Uses hybrid traces (Huawei and Azure) representing compute-heavy and I/O-heavy FaaS requests with fast-changing arrival patterns. It models cold-start latency, queuing latency, and inter-datacenter network latency.
• Utility Model: Defines utility based on request-specific latency deadlines. Utility depreciates linearly as the violation time exceeds a threshold.
• Power & Cooling Model: Estimates power consumption using a polynomial model based on CPU utilization, including IT equipment, mechanical cooling (CRAC, chillers), and internal power conditioning. It distinguishes between free-air and mechanical cooling.
• Environmental Models:
  • Energy Cost: Incorporates real-time TOU pricing and on-site solar generation.
  • Carbon Model: Accounts for carbon from electricity generation and water-related carbon (from water treatment and production).
  • Water Model: Calculates on-site evaporative water, blowdown water for wastewater treatment, and off-site water used for electricity generation.

2.2 Algorithmic Components

SOFA consists of four integrated modules (Fig. 4 in the paper):

1. Arrival Predictor: Uses a stacking-learning architecture (combining decision trees, logistic regression, and neural networks) to forecast request arrival rates and function types for the upcoming epoch. This allows for proactive container pre-warming to reduce cold-start latency.
2. Machine Learning-Guided Local Search (MLS):
   • Uses Gradient Boosting to predict the improvement potential of local search directions.
   • Employs a cumulative vector improvement metric to estimate the potential gain of a local search step without the high computational cost of calculating Pareto Hypervolume (PHV) for every step.
   • Uses K-means clustering to ensure local search efforts are distributed evenly across the design space, preventing the algorithm from getting stuck in local optima or focusing on a single region of the Pareto front.
   • Implements edge-aware PHV pruning to maintain solution diversity while removing dominated solutions.
3. Decomposition-based Evolutionary Algorithm (DEA):
   • Works in tandem with MLS to escape local optima.
   • Decomposes the solution set into "searched" and "unsearched" subsets.
   • Performs crossover and mutation between these subsets to generate new solutions outside the current neighborhood, facilitating global exploration.
4. Autoscaler:
   • Operates at the start of each epoch to adjust container sizes (vertical scaling) based on the difference between predicted and actual request arrivals.
   • Handles container over-sizing by splitting containers to fit node capacity constraints, ensuring resource efficiency and minimizing queue times.

3. Key Contributions

The paper claims the following specific contributions:

• Problem Formulation: A four-objective FaaS scheduling problem formulation that simultaneously minimizes SLA violations (utility loss), energy costs, operational carbon emissions, and water consumption in GDDC environments.
• SOFA Framework: A new framework combining stacking-learning-based workload prediction, decomposition-based evolutionary algorithms (DEA), machine-learning-guided local search (MLS), and container autoscaling.
• Comprehensive Modeling: Detailed models for FaaS workloads, computation mechanisms, inter-datacenter latency, energy consumption, utility, electricity pricing, carbon footprint (including water-related carbon), and water footprint.
• Diverse Solution Generation: The ability to generate a true Pareto front, offering cloud providers a diverse set of high-quality solutions representing different trade-offs, rather than a single fixed solution.

4. Experimental Results

The authors evaluated SOFA against five state-of-the-art frameworks: NSGA-III, SWFAGA, CASA, Greencourier (Greenc), and Kimchi. Experiments were conducted using a simulator with real-world datacenter hardware parameters and hybrid FaaS traces.

Key Findings:

• Overall Performance: Compared to the best state-of-the-art framework, SOFA achieved:
  • 32% reduction in energy costs.
  • 28% reduction in operational carbon emissions.
  • 32% reduction in water usage.
  • 3% improvement in SLA performance (reduction in utility loss).
• Solution Diversity: SOFA consistently generated a more diverse Pareto front than competitors. While other frameworks (like NSGA-III and SWFAGA) struggled to provide distinct solutions as utility constraints varied, SOFA maintained unique solutions across a wide range of constraints (2% to 10% utility loss).
• Single-Objective Optimization: When selecting the best solution for a specific objective, SOFA outperformed all others:
  • Lowest utility loss (40% lower than the second-best).
  • Lowest energy costs (16% lower than the second-best performing framework).
  • Lowest carbon emissions (1% lower).
  • Lowest water usage (82% lower than the second-best framework).
• Sensitivity Analysis:
  • Network Latency: SOFA maintained the lowest utility loss even as inter-datacenter latency increased, effectively prioritizing local execution when necessary to avoid high latency penalties.
  • Request Intensity: SOFA provided superior solutions across varying datacenter utilization rates (20% to 60%), though performance degraded at 80% utilization due to resource oversubscription, a limitation shared by other frameworks.
  • Privacy Constraints: Under strict privacy constraints (e.g., 80% of requests must stay local), SOFA still achieved the lowest energy costs, though costs increased significantly as constraints tightened.

5. Significance and Claims

The paper positions SOFA as a significant step toward SLA-aware, environmentally sustainable FaaS orchestration at a geographic scale.

• Holistic Sustainability: The authors emphasize that by integrating water usage and water-related carbon emissions into the optimization, SOFA provides a more accurate and holistic view of the environmental footprint than previous energy-centric approaches.
• Trade-off Management: The framework's ability to generate a diverse Pareto front is highlighted as crucial for cloud service providers, allowing them to make informed decisions based on specific operational priorities (e.g., prioritizing cost vs. carbon) without compromising SLAs.
• Scalability: The use of metaheuristics and machine learning guidance allows SOFA to handle the complexity of multi-objective optimization in GDDCs more efficiently than brute-force or standard genetic algorithm approaches.

The authors conclude that SOFA represents a promising algorithmic framework for achieving robust reductions in utility loss, energy costs, carbon emissions, and water usage across a wide range of geo-distributed cloud datacenter operation scenarios.