Skip to content

Benchmark

This user guide shows how to run benchmarks against a vLLM model server deployment by using both Gateway API Inference Extension, and a Kubernetes service as the load balancing strategy. The benchmark uses the Latency Profile Generator (LPG) tool to generate load and collect results.

Prerequisites

Deploy the inference extension and sample model server

Follow the getting started guide to deploy the vLLM model server, CRDs, etc.

Note: Only the GPU-based model server deployment option is supported for benchmark testing.

[Optional] Scale the sample vLLM deployment

You are more likely to see the benefits of the inference extension when there are a decent number of replicas to make the optimal routing decision.

kubectl scale deployment vllm-llama3-8b-instruct --replicas=8

Expose the model server via a k8s service

To establish a baseline, expose the vLLM deployment as a k8s service:

kubectl expose deployment vllm-llama3-8b-instruct --port=80 --target-port=8000 --type=LoadBalancer

Run benchmark

The LPG benchmark tool works by sending traffic to the specified target IP and port, and collecting the results. Follow the steps below to run a single benchmark. Multiple LPG instances can be deployed to run benchmarks in parallel against different targets.

  1. Check out the repo.

    git clone https://github.com/kubernetes-sigs/gateway-api-inference-extension
cd gateway-api-inference-extension

  2. Get the target IP. The examples below shows how to get the IP of a gateway or a k8s service.

    # Get gateway IP
GW_IP=$(kubectl get gateway/inference-gateway -o jsonpath='{.status.addresses[0].value}')
# Get LoadBalancer k8s service IP
SVC_IP=$(kubectl get service/vllm-llama2-7b -o jsonpath='{.status.loadBalancer.ingress[0].ip}')

echo $GW_IP
echo $SVC_IP

  3. Then update the <target-ip> in ./config/manifests/benchmark/benchmark.yaml to the value of $SVC_IP or $GW_IP. Feel free to adjust other parameters such as request_rates as well. For a complete list of LPG configurations, refer to the LPG user guide.

  4. Start the benchmark tool.

    kubectl apply -f ./config/manifests/benchmark/benchmark.yaml

  5. Wait for benchmark to finish and download the results. Use the benchmark_id environment variable to specify what this benchmark is for. For instance, inference-extension or k8s-svc. When the LPG tool finishes benchmarking, it will print a log line LPG_FINISHED. The script below will watch for that log line and then start downloading results.

    benchmark_id='k8s-svc' ./tools/benchmark/download-benchmark-results.bash

    After the script finishes, you should see benchmark results under ./tools/benchmark/output/default-run/k8s-svc/results/json folder. Here is a sample json file. Replace k8s-svc with inference-extension when running an inference extension benchmark.

Tips

  • When using a benchmark_id other than k8s-svc or inference-extension, the labels in ./tools/benchmark/benchmark.ipynb must be updated accordingly to analyze the results.
  • You can specify run_id="runX" environment variable when running the ./download-benchmark-results.bash script. This is useful when you run benchmarks multiple times to get a more statistically meaningful results and group the results accordingly.
  • Update the request_rates that best suit your benchmark environment.

Advanced Benchmark Configurations

Refer to the LPG user guide for a detailed list of configuration knobs.

Analyze the results

This guide shows how to run the jupyter notebook using vscode after completing k8s service and inference extension benchmarks.

  1. Create a python virtual environment.

    python3 -m venv .venv
source .venv/bin/activate

  2. Install the dependencies.

    pip install -r ./tools/benchmark/requirements.txt

  3. Open the notebook ./tools/benchmark/benchmark.ipynb, and run each cell. In the last cell update the benchmark ids withinference-extension and k8s-svc. At the end you should see a bar chart like below where "ie" represents inference extension. This chart is generated using this benchmarking tool with 6 vLLM (v1) model servers (H100 80 GB), llama2-7b and the ShareGPT dataset.

    alt text