The document covers several tricks of how you can tune your inference performance based on the engine you use including multithreading support, engine threads configuration and how to enable DNNL(MKLDNN).
One of the advantage of Deep Java Library (DJL) is Multi-threaded inference support. It can help to increase the throughput of your inference on multi-core CPUs and GPUs and reduce memory consumption compare to Python.
Predictor is not designed to be thread-safe (although some implementation is),
we recommend creating a new Predictor for each thread.
For a reference implementation, see Multi-threaded Benchmark.
you need to set corresponding configuration based on the engine you want to use.
To use Apache MXNet Engine to run multi-threading, complete the following steps.
If using the MXNet engine for a multi-threaded inference case, you need to specify the ‘MXNET_ENGINE_TYPE’ environment variable using the following command:
To get the best throughput, you may also need to set ‘OMP_NUM_THREADS’ environment variable:
Note that MxNet uses thread_local storage: Every thread that performs inference will allocate memory. In order to avoid memory leaks it is necessary to call MxNet from a fixed size thread pool. See https://github.com/apache/incubator-mxnet/issues/16431#issuecomment-562052116.
If you use multithreading inference feature with DJL 0.8.0 and earlier version, we have to disable GC to close the NDArray by
# If you are using DJL 0.5.0 -Dai.djl.pytorch.disable_close_resource_on_finalize=true # If you are using DJL 0.6.0 -Dai.djl.disable_close_resource_on_finalize=true
Please make sure all the NDArrays are attached to the NDManager.
Unlike TensorFlow and Apache MXNet, PyTorch by default doesn’t enable MKLDNN which is treated as a device type like CPU and GPU. You can enable it by
You might see the exception if certain data type or operator is not supported with the oneDNN device.
There are two configurations you can set to optimize the inference performance.
-Dai.djl.pytorch.num_interop_threads=[num of the interop threads]
It configures the number of the operations JIT interpreter fork to execute in parallel.
-Dai.djl.pytorch.num_threads=[num of the threads]
It configures the number of the threads within the operation. It is set to number of CPU cores by default.
You can find more detail in PyTorch.
When the first inference is made on a new batch size, torchscript generates an optimized execution graph for the model. This optimization can take some time and is done for each batch size you use. To disable this feature, you can use the following code:
Engine.getEngine("PyTorch"); // Make sure PyTorch engine is loaded JniUtils.setGraphExecutorOptimize(false);
By default, Graph Executor Optimization is on. It is important to note that it must be disabled for each thread that you are using the model from. If you forget to disable it on a thread, it will cause the optimization to be performed and delay any running/pending executions.
Disabling Graph Executor Optimization causes a maximum throughput and performance loss that can depend on the model and hardware. This should only be disabled when you do not have the time to “warmup” a model with your given batch sizes, or you want to use dynamic batch sizes.
You can follow the same steps as other engines for running multithreading inference using TensorFlow engine.
It’s recommended to use one
Predictor for each thread and avoid using a new
Predictor for each inference call.
You can refer to our Multithreading Benchmark as an example,
here is how to run it using TensorFlow engine.
./gradlew benchmark --args='-e TensorFlow -c 100 -t -1 -u djl://ai.djl.tensorflow/resnet/0.0.1/resnet50 -s 1,224,224,3'
By default, TensorFlow engine comes with oneDNN enabled, no special configuration needed.
It’s recommended to use 1 thread for operator parallelism during multithreading inference. You can configure it by setting the following 3 envinronment variables:
export OMP_NUM_THREADS=1 export TF_NUM_INTEROP_THREADS=1 export TF_NUM_INTRAOP_THREADS=1
DLR(TVM) itself doesn’t support multithreading. The most obvious reason is that in the implementation of forward(), it is require to setInputs, runInference followed by getOutputs. As a result, we create a new TVM model when we call newPredictor() and free the model when you call Predictor.close(). Please make sure to create a new Predictor in each thread.
TVM internally leverages full hardware resource. Based on our experiment, setting TVM_NUM_THREADS to 1 get best throughput as it avoids resource contention.