A‑LEMS: Agentic LLM Energy Measurement System¶
A cross-layer measurement and profiling framework for AI workloads.
🎯 Overview¶
A‑LEMS is a research-grade measurement and profiling framework for AI workloads. It captures telemetry across hardware, system, orchestration, and workload levels, enabling energy-aware AI systems research and evaluation of model behavior.
✨ Key Capabilities¶
| Level | Metrics Captured |
|---|---|
| Hardware | CPU package, core, uncore, DRAM energy via RAPL |
| Performance | Instructions, cycles, IPC, cache activity |
| System | Context switches, interrupts, memory faults |
| Thermal/Network | Temperature rise, inference latency |
| Workload | Prompt tokens, completion tokens, execution time |
| Orchestration | Planning, execution, synthesis phases with per-phase energy |
~144 features per run combining hardware, system, network, LLM, and orchestration metrics.
🌍 Sustainability Layer¶
Translates energy into environmental impact:
- Carbon (g CO₂) using region-aware grid factors
- Water (ml) for data center cooling
- Methane (mg CH₄) with IPCC AR6 factors
👥 Who Is This For?¶
| User | Use Case |
|---|---|
| Silicon Developers | Analyze energy/thermal behavior at hardware level |
| Orchestration Teams | Evaluate multi-agent workflow overhead |
| ML Engineers | Capture LLM telemetry for model optimization |
| Sustainability Teams | Translate energy to carbon/water/methane impact |
| Cloud Architects | Manage multi-host experiments across platforms |
🧪 Experiment Design¶
- Structured templates for systematic variation of models, tasks, workflows
- 16 configurable task categories — easily extensible
- Multi-host dispatch across multiple machines
- Per-query normalization for cross-model comparison
📊 Output & Reporting¶
After each experiment, A‑LEMS generates a detailed PDF lab report summarizing:
- Hardware-level energy breakdown
- Orchestration tax analysis
- Thermal profiles
- Sustainability metrics
- Task-level performance
🚀 Live Demos¶
Try the full-featured interface:
| Platform | URL |
|---|---|
| Streamlit | https://a-lems-dash.streamlit.app/ |
| Render | https://a-lems-dashboard.onrender.com/ |
📖 Documentation Sections¶
| Section | Description |
|---|---|
| Getting Started | Complete setup guide |
| User Guide | Running experiments |
| Developer Guide | Extending the system |
| API Reference | Technical docs |
| Research | Findings & publications |
📄 License¶
MIT License - see LICENSE file for details.
Built for energy-aware AI research