C++ Torch Server

Serve torch models as rest-api using Drogon, example included for resnet18 model for Imagenet. Benchmarks show improvement of ~6-10x throughput and latencies for resnet18 at peak load.

Build & Run Instructions

# Create Optimized models for your machine.
$ python3 optimize_model_for_inference.py

# Build and Run Server
$ docker compose run --service-ports blaze

Development

• Add Docker to CLion toolchain this will setup all necessary dependencies.

Client Instructions

curl "localhost:8088/classify" -F "image=@images/cat.jpg"

Benchmarking Instructions

# Drogon + libtorch
for i in {0..8}; do curl "localhost:8088/classify" -F "image=@images/cat.jpg"; done # Run once to warmup.
wrk -t8 -c100 -d60 -s benchmark/upload.lua "http://localhost:8088/classify" --latency

# FastAPI + pytorch
cd benchmark/python_fastapi
python3 -m venv env
source env/bin/activate
python3 -m pip install -r requirements.txt # Run just once to isntall dependencies to folder.
gunicorn main:app -w 2 -k uvicorn.workers.UvicornWorker --bind 127.0.0.1: # Best performance on my machine, tried 3/4 also.
deactivate # Use after benchmarking is done and gunicorn is closed

cd ../.. # back to root folder
for i in {0..8}; do curl "localhost:8088/classify" -F "image=@images/cat.jpg"; done
wrk -t8 -c100 -d60 -s benchmark/fastapi_upload.lua "http://localhost:8088/classify" --latency

Benchmarking results

Drogon + libtorch

# OS: Ubuntu 21.10 x86_64
# Kernel: 5.15.14-xanmod1
# CPU: AMD Ryzen 9 5900X (24) @ 3.700GHz
# GPU: NVIDIA GeForce RTX 3070
Running 1m test @ http://localhost:8088/classify
  8 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency    39.30ms   10.96ms  95.51ms   70.50%
    Req/Sec   306.58     28.78   390.00     70.92%
  Latency Distribution
     50%   37.40ms
     75%   45.69ms
     90%   54.57ms
     99%   69.34ms
  146612 requests in 1.00m, 30.34MB read
Requests/sec:   2441.60
Transfer/sec:    517.41KB

FastAPI + pytorch

# OS: Ubuntu 21.10 x86_64
# Kernel: 5.15.14-xanmod1
# CPU: AMD Ryzen 9 5900X (24) @ 3.700GHz
# GPU: NVIDIA GeForce RTX 3070
Running 1m test @ http://localhost:8088/classify
  8 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   449.50ms  239.30ms   1.64s    70.39%
    Req/Sec    33.97     26.41   121.00     83.46%
  Latency Distribution
     50%  454.64ms
     75%  570.73ms
     90%  743.54ms
     99%    1.16s
  12981 requests in 1.00m, 2.64MB read
Requests/sec:    216.13
Transfer/sec:     44.96KB

Architecture

• API request handing and model Pre-processing in the Drogon Controller controllers/ImageClass.cc
• Batched Model Inference logic & post-processing in lib/ModelBatchInference.cpp

TODOS

• Multithreaded batched inference
• FP16 Inference
• Uses c++20 coroutines for wait free event loop tasks
• Add compiler optimizations for cmake.
• Benchmark optimizations like Channel last, ONNX, TensorRT and report what's faster.
• Pin Batched tensor used for inference to memory and re-use at every inference. No Improvement.
• User Torch-TensorRT for inference, fastest on CUDA devices. Cuts down from 5ms to 1-2ms .
• Use Torch Nvjpeg for faster image decoding, currently spends 2ms on this call with libjpeg-turbo.
• Int8 Inference using FXGraph post-training quantization, Resnet Int8 Quantization example1 , example2
• Benchmark framework against mosec
• Use lockfree queues
• Seperate Pre-Process, Infer and post-preprocessing.
• Added address & memory leak sanitizers to CMake.
• Dockerize for easy usage.

Notes

• WIP: Just gets the job done for now, not production ready, though tested regularly.