AWS Deep Learning Containers Intel Math Kernel Library (MKL) Recommendations
MKL Recommendation for CPU containers
Contents
The performance for training and inference workloads for a Deep Learning framework on CPU instances can vary and depend on a variety of configuration settings. As an example, on AWS EC2 c5.18xlarge instances the number of physical cores is 36 while the number of logical cores is 72. MKL's configuration settings for training and inference are influenced by these factors. By updating MKL's configuration to match your instance's capabilities, you may achieve performance improvements.
Consider the following examples using an Intel-MKL-optimized TensorFlow binary:
-
A ResNet50v2 model, trained with TensorFlow and served for inference with TensorFlow Serving was observed to achieve 2x inference performance when the MKL settings were adjusted to match the instance's number cores. The following settings were used on a c5.18xlarge instance.
export TENSORFLOW_INTER_OP_PARALLELISM=2 # For an EC2 c5.18xlarge instance, number of logical cores = 72 export TENSORFLOW_INTRA_OP_PARALLELISM=72 # For an EC2 c5.18xlarge instance, number of physical cores = 36 export OMP_NUM_THREADS=36 export KMP_AFFINITY='granularity=fine,verbose,compact,1,0' # For an EC2 c5.18xlarge instance, number of physical cores / 4 = 36 /4 = 9 export TENSORFLOW_SESSION_PARALLELISM=9 export KMP_BLOCKTIME=1 export KMP_SETTINGS=0
-
A ResNet50_v1.5 model, trained with TensorFlow on the ImageNet dataset and using a NHWC image shape, the training throughput performance was observed to be around 9x faster. This is compared to the binary without MKL optimizations and measured in terms of samples/second. The following environment variables were used:
export TENSORFLOW_INTER_OP_PARALLELISM=0 # For an EC2 c5.18xlarge instance, number of logical cores = 72 export TENSORFLOW_INTRA_OP_PARALLELISM=72 # For an EC2 c5.18xlarge instance, number of physical cores = 36 export OMP_NUM_THREADS=36 export KMP_AFFINITY='granularity=fine,verbose,compact,1,0' # For an EC2 c5.18xlarge instance, number of physical cores / 4 = 36 /4 = 9 export KMP_BLOCKTIME=1 export KMP_SETTINGS=0
The following links will help you learn how to use to tune Intel MKL and your Deep Learning framework's settings to optimize your deep learning workload:
EC2 guide to set environment variables
Refer to
docker run documentation on how to set
environment variables when creating a
container: https://docs.docker.com/engine/reference/run/#env-environment-variables
The following is an example on setting en environment variable
called OMP_NUM_THREADS for docker run.
ubuntu@ip-172-31-95-248:~$ docker run -e OMP_NUM_THREADS=36 -it --entrypoint "" 999999999999.dkr.ecr.us-east-1.amazonaws.com/beta-tensorflow-inference:1.13-py2-cpu-build bash root@d437faf9b684:/# echo $OMP_NUM_THREADS 36
In rare cases Intel MKL can have adverse effects. To disable MKL with TensorFlow, set the following environment variables:
export TF_DISABLE_MKL=1 export TF_DISABLE_POOL_ALLOCATOR=1
ECS guide to set environment variables
To specify the environment variables for a container at runtime in ECS, you must edit the ECS task definition. Add the environment variables in the form of 'name' and 'value' key-pairs in containerDefinitions part of the task definition. The following is an example of setting OMP_NUM_THREADS and KMP_BLOCKTIME variables.
{ "requiresCompatibilities": [ "EC2" ], "containerDefinitions": [{ "command": [ "mkdir -p /test && cd /test && git clone -b r1.13 https://github.com/tensorflow/serving.git && tensorflow_model_server --port=8500 --rest_api_port=8501 --model_name=saved_model_half_plus_two_cpu --model_base_path=/test/serving/tensorflow_serving/servables/tensorflow/testdata/saved_model_half_plus_two_cpu" ], "entryPoint": [ "sh", "-c" ], "name": "EC2TFInference", "image": "999999999999.dkr.ecr.us-east-1.amazonaws.com/tf-inference:1.12-cpu-py3-ubuntu16.04", "memory": 8111, "cpu": 256, "essential": true, "environment": [{ "name": "OMP_NUM_THREADS", "value": "36" }, { "name": "KMP_BLOCKTIME", "value": 1 } ], "portMappings": [{ "hostPort": 8500, "protocol": "tcp", "containerPort": 8500 }, { "hostPort": 8501, "protocol": "tcp", "containerPort": 8501 }, { "containerPort": 80, "protocol": "tcp" } ], "logConfiguration": { "logDriver": "awslogs", "options": { "awslogs-group": "/ecs/TFInference", "awslogs-region": "us-west-2", "awslogs-stream-prefix": "ecs", "awslogs-create-group": "true" } } }], "volumes": [], "networkMode": "bridge", "placementConstraints": [], "family": "Ec2TFInference" }
In rare cases Intel MKL can have adverse effects. To disable MKL with TensorFlow, set the following environment variables:
{ "name": "TF_DISABLE_MKL", "value": 1 }, { "name": "TF_DISABLE_POOL_ALLOCATOR", "value": 1 }
EKS guide to set environment variables
To specify the environment variables for the container at runtime,
edit the raw manifests of your EKS job (.yaml, .json) . The
following snippet of a manifest shows the definition of a
container, with name squeezenet-service. Along
with other attributes such as args and ports, the environment
variables are listed in the form of 'name' and 'value' key-pairs.
containers: - name: squeezenet-service image: 999999999999.dkr.ecr.us-east-1.amazonaws.com/beta-mxnet-inference:1.4.0-py3-gpu-build command: - mxnet-model-server args: - --start - --mms-config /home/model-server/config.properties - --models squeezenet=https://s3.amazonaws.com/model-server/models/squeezenet_v1.1/squeezenet_v1.1.model ports: - name: mms containerPort: 8080 - name: mms-management containerPort: 8081 imagePullPolicy: IfNotPresent env: - name: AWS_REGION value: us-east-1 - name: OMP_NUM_THREADS value: 36 - name: TENSORFLOW_INTER_OP_PARALLELISM value: 0 - name: KMP_AFFINITY value: 'granularity=fine,verbose,compact,1,0' - name: KMP_BLOCKTIME value: 1
In rare cases Intel MKL can have adverse effects. To disable MKL with TensorFlow, set the following environment variables:
- name: TF_DISABLE_MKL value: 1 - name: TF_DISABLE_POOL_ALLOCATOR value: 1
