Benchmark your DL model¶
dll-bench is a command line tool that makes it easy for you to benchmark the model on different platforms.
With djl-bench, you can easily compare your model's behavior in different use cases, such as:
- single-threaded vs. multi-threaded
- single input vs. batched inputs
- CPU vs. GPU or other hardware accelerator
- Single GPU vs. multiple GPUs
- default engine options vs. customized engine configuration
- running with different engines
- running with different version of the engine
djl-bench currently support benchmark the following type of models:
- PyTorch TorchScript model
- TensorFlow SavedModel bundle
- Apache MXNet model
- ONNX model
- TensorRT model
- XGBoost model
- LightGBM model
- Python script model
You can build djl-bench from source if you need to benchmark fastText/BlazingText/Sentencepiece models.
Installation¶
For Ubuntu
- Install using snap
sudo snap install djlbench --classic
sudo snap alias djlbench djl-bench
- Or download .deb package from S3
curl -O https://publish.djl.ai/djl-bench/0.30.0/djl-bench_0.30.0-1_all.deb
sudo dpkg -i djl-bench_0.30.0-1_all.deb
For macOS, centOS or Amazon Linux 2
You can download djl-bench zip file from here.
curl -O https://publish.djl.ai/djl-bench/0.30.0/benchmark-0.30.0.zip
unzip benchmark-0.30.0.zip
rm benchmark-0.30.0.zip
sudo ln -s $PWD/benchmark-0.30.0/bin/benchmark /usr/bin/djl-bench
For Windows
We are considering to create a chocolatey
package for Windows. For the time being, you can
download djl-bench zip file from here.
Or you can run benchmark using gradle:
cd djl-serving
gradlew benchmark --args="--help"
Prerequisite¶
Please ensure Java 11+ is installed and you are using an OS that DJL supported with.
After that, you need to clone the djl project and cd
into the folder.
DJL supported OS:
- Ubuntu 18.04 and above
- Amazon Linux 2 and above
- MacOS latest version
- Windows 10 (Windows Server 2016+)
If you are trying to use GPU, please ensure the CUDA driver is installed. You can verify that through:
nvcc -V
to checkout the version. For different Deep Learning engine you are trying to run the benchmark, they have different CUDA version to support. Please check the individual Engine documentation to ensure your CUDA version is supported.
Sample benchmark script¶
Here is a few sample benchmark script for you to refer. You can also skip this and directly follow the 4-step instructions for your own model.
Benchmark on a Tensorflow model from tfhub url with all-zeros NDArray input for 10 times:
djl-bench -e TensorFlow -u https://tfhub.dev/tensorflow/resnet_50/classification/1 -c 10 -s 1,224,224,3
Similarly, this is for PyTorch
djl-bench -e PyTorch -u https://alpha-djl-demos.s3.amazonaws.com/model/djl-blockrunner/pytorch_resnet18.zip -n traced_resnet18 -c 10 -s 1,3,224,224
Benchmark a model from ONNX Model Zoo
djl-bench -e OnnxRuntime -u https://s3.amazonaws.com/onnx-model-zoo/resnet/resnet18v1/resnet18v1.tar.gz -s 1,3,224,224 -n resnet18v1/resnet18v1 -c 10
Benchmark from ModelZoo¶
MXNet¶
Resnet50 image classification model:
djl-bench -c 2 -s 1,3,224,224 -u djl://ai.djl.mxnet/resnet/0.0.1/resnet50_v2
PyTorch¶
SSD object detection model:
djl-bench -e PyTorch -c 2 -s 1,3,300,300 -u djl://ai.djl.pytorch/ssd/0.0.1/ssd_300_resnet50
# For AL2 or centos 7, you need use precxx11 version of PyTorch
PYTORCH_PRECXX11=true djl-bench -e PyTorch -c 2 -s 1,3,300,300 -u djl://ai.djl.pytorch/ssd/0.0.1/ssd_300_resnet50
Configuration of Benchmark script¶
To start your benchmarking, we need to make sure we provide the following information.
- The Deep Learning Engine
- The source of the model
- How many runs you would like to make
- Sample input for the model
- (Optional) Multi-thread benchmark
The benchmark script located here.
Just do the following:
djl-bench --help
This will print out the possible arguments to pass in:
usage: djl-bench [-p MODEL-PATH] -s INPUT-SHAPES [OPTIONS]
-c,--iteration <ITERATION> Number of total iterations.
-d,--duration <DURATION> Duration of the test in minutes.
-e,--engine <ENGINE-NAME> Choose an Engine for the benchmark.
-g,--gpus <NUMBER_GPUS> Number of GPUS to run multithreading inference.
-h,--help Print this help.
-l,--delay <DELAY> Delay of incremental threads.
--model-arguments <MODEL-ARGUMENTS> Specify model loading arguments.
--model-options <MODEL-OPTIONS> Specify model loading options.
-n,--model-name <MODEL-NAME> Specify model file name.
--neuron-cores <NEURON-CORES> Number of neuron cores to run multithreading inference, See
https://awsdocs-neuron.readthedocs-hosted.com.
-o,--output-dir <OUTPUT-DIR> Directory for output logs.
-p,--model-path <MODEL-PATH> Model directory file path.
-s,--input-shapes <INPUT-SHAPES> Input data shapes for the model.
-t,--threads <NUMBER_THREADS> Number of inference threads.
-u,--model-url <MODEL-URL> Model archive file URL.
-w,--warmup-iteration <WARMUP-ITERATION> Number of warmup iterations, default: 2.
--wlm Use a WorkLoad Manager benchmark
Step 1: Pick your deep engine¶
By default, the above script will use MXNet as the default Engine, but you can always change that by adding the followings:
-e TensorFlow # TensorFlow
-e PyTorch # PyTorch
-e MXNet # Apache MXNet
-e OnnxRuntime # pytorch
-e TensorRT # TensorRT
-e XGBoost # XGBoost
-e LightGBM # LightGBM
-e Python # Python script
Step 2: Identify the source of your model¶
DJL accept variety of models came from different places.
Remote location¶
Use --model-url
option to load a model from a URL. The URL must point to an archive file.
The following is a pytorch model
-u https://alpha-djl-demos.s3.amazonaws.com/model/djl-blockrunner/pytorch_resnet18.zip
We would recommend to make model files in a zip for better file tracking.
Local directory¶
Use --model-path
option to load model from a local directory or an archive file.
Mac/Linux
-p /home/ubuntu/models/pytorch_resnet18
or
-p /home/ubuntu/models/pytorch_resnet18.zip
Windows
-p C:\models\pytorch_resnet18
or
-p C:\models\pytorch_resnet18.zip
If the model file name is different from the parent folder name (or the archive file name), you need
to specify --model-name
in the --args
:
-n traced_resnet18
Step 3: Define how many runs you would like to make¶
add -c
inside with a number
-c 1000
This will run 1000 times inference.
Step 4: Define your model inputs¶
The benchmark script uses dummy NDArray inputs.
It will make fake NDArrays (like NDArray.ones
) to feed in the model for inference.
If we would like to fake an image:
-s 1,3,224,224
This will create a NDArray (DataType FLOAT32) of shape(1, 3, 224, 224).
If your model requires multiple inputs like three NDArrays with shape 1, 384 and 384. You can do the followings:
-s (1),(384),(384)
If you input DataType
is not FLOAT32, you can specify the data type with suffix:
- f: FLOAT32, this is default and is optional
- s: FLOAT16 (short float)
- d: FLOAT64 (double)
- u: UINT8 (unsigned byte)
- b: INT8 (byte)
- i: INT32 (int)
- l: INT64 (long)
- B: BOOLEAN (boolean)
For example:
-s (1)i,(384)f,(384)
Optional Step: multithreading inference¶
You can also do multi-threading inference with DJL. For example, if you would like to run the inference with 10 threads:
-t 10
Best thread number for your system: The same number of cores your system have or double of the total cores.
You can also add -l
to simulate the increment load for your inference server. It will add threads with the delay of time.
-t 10 -l 100
The above code will create 10 threads with the wait time of 100ms.
Advanced use cases¶
Benchmark for a long period of time¶
For different purposes, we designed different mode you can play with. Such as the following arg:
-d 86400
This will ask the benchmark script repeatedly running the designed task for 86400 seconds (24 hour). If you would like to make sure DJL is stable in the long run, you can do that.
Collect memory usage¶
You can also keep monitoring the DJL memory usages by enable the following flag:
export BENCHMARK_OPTS="-Dcollect-memory=true"
The memory report will be made available in build/memory.log
.
Extra model warmup iterations¶
PyTorch engine try to optimize graph execution at run time by default, this may impact the latency for the first a few inferences. You can increase warmup iterations to get more accurate benchmark result:
-w 10