RACER: Robust and Adaptive Computing and Energy Resource Coordination Framework

This repository contains the datasets and experimental results associated with the submission: Robust Restless Multi-Armed Bandit for Data Center Flexibility Services Through Virtual Machine Scheduling.

1. Data Center Dataset and Power/QoS Distribution

• datasets/datacenter_with_metrics/datacenter_*_with_metrics.csv
• power_qos_distribution/power_qos_distribution.png
• power_qos_distribution/plot_power_qos_distribution.py

Each CSV file represents one data center workload trace with VM-level resource, lifetime, ranking, power-saving, QoS-cost, and reward fields. Use the numbered files (datacenter_0_with_metrics.csv through datacenter_19_with_metrics.csv) as independent data center instances for experiments or comparisons.

Regenerate:

python power_qos_distribution/plot_power_qos_distribution.py

2. ST, TW, and TM-TW Performance Comparison

This experiment compares State Thompson (ST), Thompson-Whittle (TW), and Trust-mixed Thompson-Whittle (TM-TW) policies for adaptive scheduling across the data center instances.

Files are available in st_tw_tmtw_comparison/.

3. TM-TW Ablation Experiment

This ablation evaluates the contribution of the trust-mixing component in TM-TW by comparing reward outcomes against the EXP4 strategy and related variants.

Files are available in tmtw_ablation/.

Regenerate the two comparison plots:

python tmtw_ablation/replot_exp4_vs_ablation_reward.py

4. Sensitivity Experiments

These experiments test how performance changes under different workload sizes and contextual noise levels.

Files are available in sensitivity/.

Regenerate:

python sensitivity/replot_n_jobs.py
python sensitivity/replot_contextual_noise.py