Triton Performance Analyzer

Triton Performance Analyzer is CLI tool which can help you optimize the inference performance of models running on Triton Inference Server by measuring changes in performance as you experiment with different optimization strategies.

⚠️ Warning: genai-perf is being deprecated. Please migrate to AIPerf for continued support and enhanced features.

Features

Inference Load Modes

• Concurrency Mode simlulates load by maintaining a specific concurrency of outgoing requests to the server

• Request Rate Mode simulates load by sending consecutive requests at a specific rate to the server

• Custom Interval Mode simulates load by sending consecutive requests at specific intervals to the server

Performance Measurement Modes

• Time Windows Mode measures model performance repeatedly over a specific time interval until performance has stabilized

• Count Windows Mode measures model performance repeatedly over a specific number of requests until performance has stabilized

Other Features

• Sequence Models, Ensemble Models, and Decoupled Models can be profiled in addition to standard/stateless/coupled models

• Input Data to model inferences can be auto-generated or specified as well as verifying output

• TorchServe can be used as the inference server in addition to the default Triton server

Quick Start

The steps below will guide you on how to start using Perf Analyzer.

Step 1: Start Triton Container

export RELEASE=<yy.mm> # e.g. to use the release from the end of February of 2023, do `export RELEASE=23.02`

docker pull nvcr.io/nvidia/tritonserver:${RELEASE}-py3

docker run --gpus all --rm -it --net host nvcr.io/nvidia/tritonserver:${RELEASE}-py3

Step 2: Download simple Model

# inside triton container
git clone --depth 1 https://github.com/triton-inference-server/server

mkdir model_repository ; cp -r server/docs/examples/model_repository/simple model_repository

Step 3: Start Triton Server

# inside triton container
tritonserver --model-repository $(pwd)/model_repository &> server.log &

# confirm server is ready, look for 'HTTP/1.1 200 OK'
curl -v localhost:8000/v2/health/ready

# detach (CTRL-p CTRL-q)

Step 4: Start Triton SDK Container

docker pull nvcr.io/nvidia/tritonserver:${RELEASE}-py3-sdk

docker run --gpus all --rm -it --net host nvcr.io/nvidia/tritonserver:${RELEASE}-py3-sdk

Step 5: Run Perf Analyzer

# inside sdk container
perf_analyzer -m simple

See the full quick start guide for additional tips on how to analyze output.

Documentation

• Installation
• Perf Analyzer CLI
• Inference Load Modes
• Input Data
• Measurements & Metrics
• Benchmarking

Contributing

Contributions to Triton Perf Analyzer are more than welcome. To contribute please review the contribution guidelines, then fork and create a pull request.

Reporting problems, asking questions

We appreciate any feedback, questions or bug reporting regarding this project. When help with code is needed, follow the process outlined in the Stack Overflow (https://stackoverflow.com/help/mcve) document. Ensure posted examples are:

• minimal - use as little code as possible that still produces the same problem

• complete - provide all parts needed to reproduce the problem. Check if you can strip external dependency and still show the problem. The less time we spend on reproducing problems the more time we have to fix it

• verifiable - test the code you're about to provide to make sure it reproduces the problem. Remove all other problems that are not related to your request/question.