ChatQnA Benchmarking

This folder contains a collection of Kubernetes manifest files for deploying the ChatQnA service across scalable nodes. It includes a comprehensive benchmarking tool that enables throughput analysis to assess inference performance.

By following this guide, you can run benchmarks on your deployment and share the results with the OPEA community.

Purpose

We aim to run these benchmarks and share them with the OPEA community for three primary reasons:

  • To offer insights on inference throughput in real-world scenarios, helping you choose the best service or deployment for your needs.

  • To establish a baseline for validating optimization solutions across different implementations, providing clear guidance on which methods are most effective for your use case.

  • To inspire the community to build upon our benchmarks, allowing us to better quantify new solutions in conjunction with current leading llms, serving frameworks etc.

Metrics

The benchmark will report the below metrics, including:

  • Number of Concurrent Requests

  • End-to-End Latency: P50, P90, P99 (in milliseconds)

  • End-to-End First Token Latency: P50, P90, P99 (in milliseconds)

  • Average Next Token Latency (in milliseconds)

  • Average Token Latency (in milliseconds)

  • Requests Per Second (RPS)

  • Output Tokens Per Second

  • Input Tokens Per Second

Results will be displayed in the terminal and saved as CSV file named 1_stats.csv for easy export to spreadsheets.

Table of Contents

Deployment

Prerequisites

  • Kubernetes installation: Use kubespray or other official Kubernetes installation guides.

  • Helm installation: Follow the Helm documentation to install Helm.

  • Setup Hugging Face Token

    To access models and APIs from Hugging Face, set your token as environment variable.

    export HF_TOKEN="insert-your-huggingface-token-here"

  • Prepare Shared Models (Optional but Strongly Recommended)

    Downloading models simultaneously to multiple nodes in your cluster can overload resources such as network bandwidth, memory and storage. To prevent resource exhaustion, it’s recommended to preload the models in advance.

    pip install -U "huggingface_hub[cli]"
sudo mkdir -p /mnt/models
sudo chmod 777 /mnt/models
huggingface-cli download --cache-dir /mnt/models Intel/neural-chat-7b-v3-3
export MODEL_DIR=/mnt/models

    Once the models are downloaded, you can consider the following methods for sharing them across nodes:

    • Persistent Volume Claim (PVC): This is the recommended approach for production setups. For more details on using PVC, refer to PVC.

    • Local Host Path: For simpler testing, ensure that each node involved in the deployment follows the steps above to locally prepare the models. After preparing the models, use --set global.modelUseHostPath=${MODELDIR} in the deployment command.

  • Add OPEA Helm Repository:

    python deploy.py --add-repo

  • Label Nodes

    python deploy.py --add-label --num-nodes 2

Deployment Scenarios

The example below are based on a two-node setup. You can adjust the number of nodes by using the --num-nodes option.

By default, these commands use the default namespace. To specify a different namespace, use the --namespace flag with deploy, uninstall, and kubernetes command. Additionally, update the namespace field in benchmark.yaml before running the benchmark test.

For additional configuration options, run python deploy.py --help

Case 1: Baseline Deployment with Rerank

Deploy Command (with node number, Hugging Face token, model directory specified):

python deploy.py --hf-token $HF_TOKEN --model-dir $MODEL_DIR --num-nodes 2 --with-rerank

Uninstall Command:

python deploy.py --uninstall

Case 2: Baseline Deployment without Rerank

python deploy.py --hf-token $HFTOKEN --model-dir $MODELDIR --num-nodes 2

Case 3: Tuned Deployment with Rerank

python deploy.py --hf-token $HFTOKEN --model-dir $MODELDIR --num-nodes 2 --with-rerank --tuned

Benchmark

Test Configurations

Key

Value

Workload

ChatQnA

Tag

V1.1

Models configuration

Key

Value

Embedding

BAAI/bge-base-en-v1.5

Reranking

BAAI/bge-reranker-base

Inference

Intel/neural-chat-7b-v3-3

Benchmark parameters

Key

Value

LLM input tokens

1024

LLM output tokens

128

Number of test requests for different scheduled node number:

Node count

Concurrency

Query number

1

128

640

2

256

1280

4

512

2560

More detailed configuration can be found in configuration file benchmark.yaml.

Test Steps

Use kubectl get pods to confirm that all pods are READY before starting the test.

Upload Retrieval File

Before testing, upload a specified file to make sure the llm input have the token length of 1k.

Get files:

wget https://github.com/opea-project/GenAIEval/tree/main/evals/benchmark/data/upload_file_no_rerank.txt
wget https://github.com/opea-project/GenAIEval/tree/main/evals/benchmark/data/upload_file.txt

Retrieve the ClusterIP of the chatqna-data-prep service.

kubectl get svc

Expected output:

chatqna-data-prep         ClusterIP   xx.xx.xx.xx    <none>        6007/TCP            51m

Use the following cURL command to upload file:

cd GenAIEval/evals/benchmark/data
# RAG with Rerank
curl -X POST "http://${cluster_ip}:6007/v1/dataprep" \
     -H "Content-Type: multipart/form-data" \
     -F "files=@./upload_file.txt"
# RAG without Rerank
curl -X POST "http://${cluster_ip}:6007/v1/dataprep" \
     -H "Content-Type: multipart/form-data" \
     -F "files=@./upload_file_no_rerank.txt"

Run Benchmark Test

Run the benchmark test using:

bash benchmark.sh -n 2

The -n argument specifies the number of test nodes. Required dependencies will be automatically installed when running the benchmark for the first time.

Data collection

All the test results will come to the folder GenAIEval/evals/benchmark/benchmark_output.

Teardown

After completing the benchmark, use the following commands to clean up the environment:

Remove Node Labels:

python deploy.py --delete-label

Delete the OPEA Helm Repository:

python deploy.py --delete-repo