GPU Cluster Capacity Simulator

A lightweight simulator for GPU cluster capacity planning for AI workloads.
Models utilization, queueing delay, packing / fragmentation, and a simplified cost view.

This project is intentionally infrastructure-focused.
It demonstrates how AI compute systems behave under load, rather than how to train or optimize ML models.

---

Demo

Open the Live Streamlit Demo

---

What this simulator models

Given a cluster configuration and workload arrival patterns, the simulator estimates:

• GPU utilization over time
• Queue depth and wait time (including P50 / P95 indicators)
• Inference SLA violations
• Cost efficiency (GPU-hours with overhead)

This mirrors the real tradeoffs faced in AI infrastructure, platform, and MLOps environments.

---

Architecture

Execution flow

1. Workload generator creates training and inference jobs
2. Scheduler assigns jobs to GPUs (capturing packing + fragmentation effects)
3. Simulator advances time and tracks system state
4. Metrics and reports are generated for analysis

---

Quickstart

1) Install

python -m venv .venv
source .venv/bin/activate
python -m pip install -U pip
pip install -e ".[dev]"

2) Run a simulation

gpu-sim run --config configs/base.yaml \
  --out results/base_tick.csv \
  --out-jobs results/base_jobs.csv

Outputs

• results/*_tick.csv → utilization, queue depth, running/completed jobs
• results/*_jobs.csv → per-job wait time, assignment, SLA violations

---

Scenarios Included

Config	Purpose
configs/base.yaml	Balanced training + inference baseline
configs/mixed_workloads.yaml	Contention and fragmentation under mixed loads
configs/inference_spike.yaml	SLA pressure during inference bursts
configs/nvl_vs_pcie.yaml	Sensitivity to reduced interconnect efficiency

Run all scenarios:

gpu-sim run --config configs/base.yaml --out results/base_tick.csv --out-jobs results/base_jobs.csv
gpu-sim run --config configs/mixed_workloads.yaml --out results/mixed_tick.csv --out-jobs results/mixed_jobs.csv
gpu-sim run --config configs/inference_spike.yaml --out results/spike_tick.csv --out-jobs results/spike_jobs.csv
gpu-sim run --config configs/nvl_vs_pcie.yaml --out results/pcie_tick.csv --out-jobs results/pcie_jobs.csv

---

Results & Insights

GPU Utilization

Insight

• High average utilization does not guarantee low latency
• Mixed workloads reduce effective throughput due to fragmentation

---

Queue Depth & SLA Pressure

Insight

• Queue depth grows rapidly under bursty inference traffic
• SLA violations often appear before GPUs look fully saturated

---

Streamlit Dashboard (Live, Interactive)

Run an interactive dashboard to compare scenarios and visualize utilization and queueing behavior.

pip install -r requirements.txt
streamlit run app.py

The dashboard supports:

• Scenario selection
• Live simulation execution
• KPI cards (utilization, P95 wait, SLA rate, cost)
• Time-series plots for utilization and queue depth

---

Development & Quality

pytest -q
ruff check .

CI runs automatically via GitHub Actions.

---

Why this project exists

AI infrastructure performance is rarely a pure “GPU count” problem.

This simulator highlights why real-world outcomes depend on:

• Workload mix (training vs inference)
• Scheduling and packing
• Queueing dynamics
• SLA design
• Cost vs reliability tradeoffs

These considerations are central to modern AI platforms, cloud infrastructure, and MLOps systems.

---

Roadmap

• Pool partitioning (training vs inference)
• Priority scheduling and preemption
• Trace-driven workloads
• Multi-cluster comparisons
• More detailed cost models (power, networking)

---