PMEC-X: Predictive Multifactor Energy Consolidation Scheduler
PMEC-X is a heterogeneous cloud task consolidation algorithm designed to reduce idle power waste in data centers while preserving SLA compliance. The scheduler combines next-epoch workload prediction, urgency-based task ordering, 7-factor VM scoring, and SLA-tier-aware consolidation into a single scheduling loop that runs every 120 seconds.
Cloud data centers consume a significant share of global electricity, and a large part of that draw comes from underutilised virtual machines. PMEC-X addresses this by combining EWMA+CUSUM workload prediction with carbon-aware and thermal-aware scheduling, then applying a consolidation policy that can retain, migrate, queue, or shut down VMs depending on SLA urgency. In the 24-hour simulation used in this repository, PMEC-X reduced energy use to 74.60 kWh, achieved zero SLA violations, and detected 30 regime changes in one epoch each.
| Scheduler | Energy (kWh) | Violations | Viol % | Avg VMs | Shutdowns |
|---|---|---|---|---|---|
| PMEC-X | 74.60 | 0 | 0.0% | 16.7 | 4 |
| MinMin | 87.53 | 0 | 0.0% | 20.0 | 0 |
| Round Robin | 125.10 | 294 | 7.1% | 20.0 | 0 |
| FCFS | 119.88 | 10 | 0.2% | 20.0 | 0 |
| MaxMin | 121.44 | 187 | 4.5% | 20.0 | 0 |
PMEC-X saves 14.8% energy vs MinMin and 40.4% vs Round Robin while maintaining zero SLA violations.
| # | Contribution | Description |
|---|---|---|
| C1 | Formal problem definition | Heterogeneous cloud task consolidation with seven normalised scoring factors, including carbon intensity and server thermal state |
| C2 | EWMA+CUSUM hybrid prediction | Next-epoch VM load forecasting with a theoretical MSE bound and regime-change detection |
| C3 | SLA-tier-aware consolidation | Differentiates CRITICAL, STANDARD, and BULK tasks to enable aggressive consolidation without SLA degradation |
| C4 | Comparative simulation framework | Benchmarks PMEC-X against Round Robin, MinMin, FCFS, and MaxMin on a 4,150-task, 24-hour workload |
PMEC-X runs a 4-phase scheduling loop every 120 seconds:
INPUT: Task Queue → VM Pool → Carbon + Thermal Signals
│
▼
PHASE 1: EWMA + CUSUM Prediction
Forecast next-epoch VM load and detect regime changes
│
▼
PHASE 2: Urgency Sort
Rank tasks by priority and deadline slack
U(τ) = 0.6 × Priority + 0.4 × Deadline Slack
│
▼
PHASE 3: 7-Factor Scoring Engine
S(τ, v) = Σ wₖ · fₖ for every (task, VM) pair
Speed · Headroom · Cost · Energy · Carbon · Thermal · Urgency
│
▼
PHASE 4: SLA-Tier-Aware Consolidation
VM load < 12% → RETAIN / MIGRATE / QUEUE / SHUTDOWN
│
▼
OUTPUT: Schedule · Energy Log · SLA Report · CUSUM Events
PMECX/
├── core/
│ ├── models.py # Task, VM, Container, Host data models
│ ├── predictor.py # EWMA + CUSUM hybrid workload predictor
│ └── scorer.py # 7-factor weighted scoring engine
│
├── simulation/
│ ├── workload.py # Poisson workload generator (4,150 tasks/24hr)
│ └── environment.py # Full scheduling loop + all baseline schedulers
│
├── baselines/
│ ├── roundrobin.py # Round Robin baseline
│ ├── fcfs.py # First Come First Served baseline
│ ├── minmin.py # MinMin baseline
│ └── maxmin.py # MaxMin baseline
│
├── dashboard/
│ └── app.py # Live Streamlit dashboard
│
├── results/ # Simulation outputs (CSV)
└── run_simulation.py # One-command runner
Prerequisites: Python 3.10+, Git
# Clone the repository
git clone https://github.com/varun222004/PMECx.git
cd PMECx
# Create and activate virtual environment
python -m venv venv
# Windows
venv\Scripts\activate
# Mac / Linux
source venv/bin/activate
# Install dependencies
pip install streamlit plotly matplotlib pandas numpypython run_simulation.pyExpected output:
PMEC-X 74.60 kWh 0.0% 16.7 VMs 4 shutdowns
Round Robin 125.10 kWh 7.1% 20.0 VMs 0 shutdowns
MinMin 87.53 kWh 0.0% 20.0 VMs 0 shutdowns
FCFS 119.88 kWh 0.2% 20.0 VMs 0 shutdowns
Energy saved: +14.8% vs MinMin | +40.4% vs Round Robin
CUSUM regime changes: 30
streamlit run dashboard/app.pyOpens at http://localhost:8501. Press ▶ Start to watch all 4 schedulers run simultaneously in real time.
| Feature | Description |
|---|---|
| Real-time energy chart | All 4 schedulers plotted live — watch PMEC-X stay low |
| VM heatmap | Live load per VM — see consolidation happening |
| CUSUM alarm box | Turns red when a regime change is detected |
| Carbon intensity dial | Shows the per-rack carbon signal used by the scorer |
| Weight sliders | Adjust all 7 scoring weights live |
| Spike injection | Simulate a sudden traffic burst — watch CUSUM fire |
| Results summary | Final per-scheduler metrics at the end of the run |
EWMA: û(t+1) = 0.3 · u(t) + 0.7 · û(t)
MSE bound: 0.053 · σ²
CUSUM: C_high(t+1) = max(0, C_high(t) + deviation − k)
Alarm when C_high > h = 0.3
- EWMA tracks smooth load in stable regimes
- CUSUM detects sudden spikes and fires in 1 epoch vs 5–8 for EWMA alone
- Detected 30 regime changes across the 24-hour simulation trace
S(τᵢ, vⱼ) = w₁·f_speed + w₂·f_headroom + w₃·f_cost + w₄·f_energy
+ w₅·f_carbon + w₆·f_thermal + w₇·f_urgency
All factors normalised to [0, 1] · Equal weights (1/7) by default
VM load < 12% ?
├── Task = CRITICAL or STANDARD → RETAIN (never move)
├── Task = BULK + slack ≥ migration cost → MIGRATE (live migration)
├── Task = BULK + slack ≥ 0 → QUEUE (defer safely)
└── VM now empty → SHUTDOWN (idle power eliminated)
| Phase | Operation | Cost |
|---|---|---|
| Prediction | EWMA + CUSUM per VM | O(m) |
| Urgency sort | Quicksort | O(n log n) |
| Assignment | n tasks × m VMs | O(n·m) |
| Consolidation | Per VM per task | O(n + m) |
| Total | Dominated by assignment | O(n·m + n log n) |
Completes each epoch in < 1ms for n=4,150 tasks, m=20 VMs.
| Parameter | Value |
|---|---|
| VM pool | 20 VMs across 4 types (fast/ standard/ cheap/ hot-aisle) |
| MIPS ranges | fast: 8,000–10,000; standard: 3,000–6,000; cheap: 800–2,000; hot-aisle: 4,000–7,000 |
| Task count | 4,150 tasks over 24 hours |
| Arrival model | Non-homogeneous Poisson (λ = 3–18 tasks/epoch) |
| SLA tiers | 10% CRITICAL · 30% STANDARD · 60% BULK |
| Epoch duration | 120 seconds · 720 epochs = 24 hours |
| EWMA α | 0.3 |
| CUSUM k / h | 0.05 / 0.3 |
| Consolidation threshold | 0.12 (θ_low) |
| Carbon scores | Rack A 0.25 · Rack B 0.50 · Rack C 0.45 · Rack D 0.75 |
| Random seed | 42 (fully reproducible) |
- SDG 7 — Affordable and Clean Energy: reduces data center electricity consumption and carbon emissions
- SDG 9 — Industry, Innovation and Infrastructure: novel algorithm contribution to sustainable cloud computing
- Live Carbon API Integration — real-time gCO₂/kWh signals via WattTime or Electricity Maps
- Bayesian Adaptive Weight Tuning — Gaussian Process bandit optimisation of the weight vector W
- Multi-Datacenter Scheduling — add network latency as an eighth scoring factor
- CloudSim Plus Validation — benchmark against Google Cluster Trace 2019
- Kubernetes Integration — production deployment of the container-layer scheduling component
- Beloglazov, A., Abawajy, J., & Buyya, R. — Energy-aware resource allocation heuristics, FGCS 2012
- Braun, T. D. et al. — Comparison of eleven static heuristics for heterogeneous task mapping, JPDC 2001
- Fan, X., Weber, W. D., & Barroso, L. A. — Power provisioning for a warehouse-sized computer, ISCA 2007
- Calheiros, R. N. et al. — CloudSim: A toolkit for cloud computing simulation, SPE 2011
- Page, E. S. — Continuous inspection schemes, Biometrika 1954
- Xu, M., Tian, W., & Buyya, R. — Survey on load balancing algorithms for VM placement, C&C 2017
Authors
- Varun V — Dept. of Information Technology, Alliance University
- Valmeeki Singh — Dept. of Information Technology, Alliance University
- Santhosh G — Dept. of Information Technology, Alliance University
- Abdul Wahab — Dept. of Information Technology, Alliance University
Repository maintainer: Varun
⚡ PMEC-X — Predict. Consolidate. Save.
Built from scratch in Python. No cloud account needed. Runs entirely on your laptop.