This 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 compared to Python.
The DJL Predictor
is not designed to be thread-safe.
Because some engines are not thread-safe, we use the Predictor to help cover the differences.
For engines that are thread-safe, we do nothing and for ones that are not, we will make copies as necessary.
Therefore, we recommend creating a new Predictor for each thread. That Predictor should be reused if the thread does multiple predictions. Alternatively, you can also use a pool of Predictors or you can leverage the DJL Serving WorkLoadManager.
For a reference implementation, see Multi-threaded Benchmark.
In addition, you may need to set engine-specific configurations as well. Engine-specific details are given below. You can also reference the list of all DJL system configurations.
To use Apache MXNet Engine to run multi-threading, complete the following steps:
By default, MXNet tries to execute operators lazily and with internal multi-threading designed to occupy the entire machine. This works well during Training and if you want to focus on latency over throughput. If you are using the MXNet engine for a Java multi-threaded inference case, you need to specify the ‘MXNET_ENGINE_TYPE’ environment variable using the following command:
export MXNET_ENGINE_TYPE=NaiveEngine
To get the best throughput, you may also need to set the ‘OMP_NUM_THREADS’ environment variable:
export OMP_NUM_THREADS=1
The PyTorch graph executor optimizer (JIT tensorexpr fuser) is enabled by default. When the first a few inferences is made on a new batch size, torchscript generates an optimized execution graph for the model.
Disabling Graph Executor Optimization causes a maximum throughput and performance loss that can depend on the model and hardware. Use djl-bench to check the impact of this optimization on your 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 disable graph executor optimizer globally by setting the following system properties:
-Dai.djl.pytorch.graph_optimizer=false
The graph executor optimizer is a per thread configuration. 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.
Engine.getEngine("PyTorch"); // Make sure PyTorch engine is loaded
JniUtils.setGraphExecutorOptimize(false);
Unlike TensorFlow and Apache MXNet, PyTorch by default doesn’t enable MKLDNN by default. Instead, it is treated as a device type like CPU and GPU. You can enable it by setting the environment variable:
-Dai.djl.pytorch.use_mkldnn=true
You might see an exception if a data type or operator is not supported with the oneDNN device.
AWS Graviton3(E) (e.g. c7g/m7g/r7g, c7gn and Hpc7g instances) supports BF16 format for ML acceleration. This can be enabled in oneDNN by setting the below environment variable
grep -q bf16 /proc/cpuinfo && export DNNL_DEFAULT_FPMATH_MODE=BF16
To avoid redundant primitive creation latency overhead, enable primitive caching by setting the LRU cache capacity. Please note this caching feature increases the memory footprint. It is recommended to tune the capacity to an optimal value for a given use case.
export LRU_CACHE_CAPACITY=1024
In addition to avoiding the redundant allocations, tensor memory allocation latencies can be optimized with Linux transparent huge pages (THP). To enable THP allocations, set the following torch environment variable.
export THP_MEM_ALLOC_ENABLE=1
Please refer to PyTorch Graviton tutorial for more details on how to achieve the best PyTorch inference performance on AWS Graviton3 instances.
PyTorch has a special flag that is used for a CNN or related network speed up. If your input size won’t change frequently, you may benefit from enabling this configuration in your model:
-Dai.djl.pytorch.cudnn_benchmark=true
If your input shape changes frequently, this change may stall your performance. For more information, check this article.
There are two configurations you can set to optimize the inference performance.
-Dai.djl.pytorch.num_interop_threads=[num of the interop threads]
This configures the number of operations the JIT interpreter forks will execute in parallel.
-Dai.djl.pytorch.num_threads=[num of the threads]
This configures the number of threads within the operation. It is set to the number of CPU cores by default.
You can find more details in PyTorch.
You can follow the same steps as for other engines for running multithreading inference using the 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.
cd djl-serving
./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, the 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 environment variables:
export OMP_NUM_THREADS=1
export TF_NUM_INTEROP_THREADS=1
export TF_NUM_INTRAOP_THREADS=1
You can use the following settings for thread optimization in Criteria
Criteria.builder()
.optOption("interOpNumThreads", <num_of_thread>)
.optOption("intraOpNumThreads", <num_of_thread>)
...
Tips: Set to 1 on both of them at the beginning to see the performance.
Then, set to total_cores
/total_java_inference_thread
on one of them to see how performance goes.
If you have tensorRT installed, you can try with the following backend on ONNXRuntime for performance optimization in Criteria
Criteria.builder()
.optOption("ortDevice", "TensorRT")
...