Master Ollama Deployment: Optimize Environment Variables for Peak Performance

In Ollama's local deployment and performance tuning, environment variables act as the "central nervous system," allowing developers to finely control model runtime behavior across single‑machine, cluster, and edge scenarios.

Cross‑Platform Environment Variable Guide

Linux/macOS Configuration

Temporary (single session)

# Quick start with custom config
export OLLAMA_PORT=12345 # custom service port
export OLLAMA_MODEL_DIR=./custom-models # dedicated model storage path
ollama serve --listen :$OLLAMA_PORT # load env vars at startup

Permanent (global)

Edit the appropriate shell config (example for ZSH):

echo 'export OLLAMA_NUM_GPUS=1' >> ~/.zshrc
echo 'export OLLAMA_CACHE_DIR="/data/ollama-cache"' >> ~/.zshrc
source ~/.zshrc # apply changes immediately

Windows GUI Configuration

Open Control Panel → System → Advanced system settings.

In "Environment Variables" add a new system variable:

Validate the configuration via the command line:

echo $env:OLLAMA_MODEL_DIR # verify custom path

Docker Container Deployment

# Dockerfile example
FROM ollama/ollama:latest
ENV OLLAMA_PORT=11434 \
    OLLAMA_USE_MLOCK=1
VOLUME /ollama/models # persist model files

Run with dynamic injection:

docker run -d \
  -p 11434:11434 \
  -v $(pwd)/models:/ollama/models \
  -e OLLAMA_GPU_LAYERS=32 \
  ollama/ollama:latest

GPU Resource Utilization Strategies

Ample VRAM (≥16GB)

export OLLAMA_ENABLE_CUDA=1
export OLLAMA_GPU_LAYERS=40
export OLLAMA_USE_MLOCK=1

Monitor with nvidia-smi and ensure GPU-Util stays above 80%.

Limited VRAM (≤8GB)

export OLLAMA_GPU_LAYERS=20
export OLLAMA_MAX_GPU_MEMORY=6GB
export OLLAMA_ENABLE_CUDA=1

Pair with nvtop to avoid OOM errors.

Concurrent Performance Optimization

High‑Concurrency API Service

export OLLAMA_MAX_WORKERS=8
export OLLAMA_NUM_THREADS=16
export OLLAMA_CACHE_SIZE=8GB
export OLLAMA_KEEP_ALIVE_TIMEOUT=60s

QPS can increase by 30‑50%, suitable for e‑commerce or chatbot workloads.

Lightweight Deployment (Laptop/Edge)

export OLLAMA_MAX_WORKERS=2
export OLLAMA_NUM_THREADS=4
export OLLAMA_CACHE_SIZE=2GB

Ideal for local knowledge‑base queries or single‑user code assistance.

Production‑Grade Security Hardening

API Access Control

# Basic auth + HTTPS encryption
export OLLAMA_AUTH_TOKEN="$(openssl rand -hex 32)"
export OLLAMA_ALLOW_ORIGINS="https://api.yourdomain.com"
export OLLAMA_ENABLE_TLS=1
export OLLAMA_TLS_CERT_FILE="/ssl/cert.pem"

Data Security Policies

# Prevent remote model pulls and enable read‑only mode
export OLLAMA_DISABLE_REMOTE_PULL=1
export OLLAMA_READ_ONLY=1
export OLLAMA_ENABLE_SANDBOX=1

Security Monitoring

# Logging and request throttling
export OLLAMA_LOG_LEVEL=INFO
export OLLAMA_LOG_FILE="/var/log/ollama/access.log"
export OLLAMA_MAX_REQUEST_SIZE=10MB

Advanced Configuration & Source‑Level Tuning

By reading Ollama's source ( envconfig/config.go), you can unlock hidden options such as:

export OLLAMA_FLASH_ATTENTION=1 # enable FlashAttention for long‑text inference
export OLLAMA_LLM_LIBRARY=llama.cpp # force specific inference library
export OLLAMA_MAX_LOADED_MODELS=3 # load up to 3 models simultaneously

Common Troubleshooting

Issue

Possible Cause

Solution

Port conflict

Multiple instances using same port

Change OLLAMA_PORT=11435 and restart

Model load failure

Insufficient directory permissions

Ensure OLLAMA_MODEL_DIR is readable/writable

GPU usage < 50%

CUDA not enabled or low layer count

Set OLLAMA_ENABLE_CUDA=1 and increase GPU_LAYERS No relevant logs

Log level too high

Set

OLLAMA_LOG_LEVEL=DEBUG

Appendix: Frequently Used Ollama Environment Variables

Key variables include OLLAMA_NUM_GPUS, OLLAMA_GPU_LAYERS, OLLAMA_ENABLE_CUDA, OLLAMA_USE_MLOCK, OLLAMA_AUTH_TOKEN, OLLAMA_ENABLE_TLS, and many others for model management, performance, and security.

After configuring, verify the setup with curl http://localhost:11434/api/status to monitor model loading and resource usage, ensuring the configuration meets expectations and delivers high‑performance, secure AI services.