TensorFlow at TACC
Last update: June 08, 2020

Scientists across domains are actively exploring and adopting deep learning as a cutting-edge methodology to make research breakthrough. At TACC, our mission is to enable discoveries that advance science and society through the application of advanced computing technologies. Thus, we are embracing this new type of application on our high end computing platforms.

TACC supports the TensorFlow+Horovod stack. This framework exposes high level interfaces for deep learning architecture specification, model training, tuning, and validation. Deep learning practitioners and domain scientists who are exploring the deep learning methodology should consider this framework for their research.

This document details how to install TensorFlow, then download and run benchmarks in both single- and multi-node modes. Due to variations in TensorFlow and Python versions, and their compatabilities with the Intel compilers and CUDA libraries, the installation instructions are quite specific. Pay careful attention to the installation instructions.

Installations at TACC

TensorFlow is installed on TACC's Frontera, Stampede2, Longhorn and Maverick2 resources.

  • Parallel Training with TensorFlow and Horovod is available on both Stampede2 and Maverick2.
  • TensorFlow v2.1 is available on Stampede2.
  • Current Longhorn Tensorflow installations are 1.13.1, 1.14.0, 1.15.2, 2.1.0

Before you begin, note that all of the following examples are run on compute, not login, nodes.
Running programs or doing computations on the login nodes may result in account suspension.
Use TACC's idev utility to grab compute node/s when conducting any TensorFlow activities.

TensorFlow on Maverick2

These instructions detail installing and running TensorFlow benchmarks on Maverick2. Maverick2 runs TensorFlow 2.1.0 with Python 3.7.0 and Intel 18.

Maverick2 supports CUDA/10.0, and CUDA/10.1. Use the respective CUDA version for your TensorFlow installation with Python 3.7.

c123-456$ module load intel/18.0.2 python3/3.7.0
c123-456$ module load cuda/10.1 cudnn/7.6.5 nccl/2.5.6
c123-456$ pip3 install --user grpcio==1.28.1 tensorflow-gpu==2.1.0 --no-cache-dir

We suggest installing Horovod version 0.19.2. If you wish to install other versions of Horovod, please submit a support ticket with the subject "Request for Horovod" and TACC staff will provide special instructions.

c123-456$ HOROVOD_CUDA_HOME=$TACC_CUDA_DIR HOROVOD_NCCL_HOME=$TACC_NCCL_DIR CC=gcc \
    HOROVOD_GPU_ALLREDUCE=NCCL HOROVOD_GPU_BROADCAST=NCCL HOROVOD_WITH_TENSORFLOW=1 pip3 install \
    --user horovod==0.19.2 --no-cache-dir

Single-Node

Download the tensorflow benchmark to your $WORK directory, then check out the branch that matches your tensorflow version.

c123-456$ cdw; git clone https://github.com/tensorflow/benchmarks.git
c123-456$ cd benchmarks
c123-456$ git checkout cnn_tf_v2.1_compatible

Benchmark the performance with the synthetic dataset on 1 GPU:

c123-456$ cd scripts/tf_cnn_benchmarks
c123-456$ module load intel/18.0.2 python3/3.7.0 cuda/10.1 cudnn/7.6.5 nccl/2.5.6
c123-456$ python3 tf_cnn_benchmarks.py --num_gpus=1 --model resnet50 --batch_size 32 --num_batches 200

Benchmark the performance with the synthetic dataset on 4 GPUs:

c123-456$ cd scripts/tf_cnn_benchmarks
c123-456$ module load intel/18.0.2 python3/3.7.0 cuda/10.1 cudnn/7.6.5 nccl/2.5.6
c123-456$ ibrun -np 4 python3 tf_cnn_benchmarks.py --variable_update=horovod \
            --num_gpus=1 --model resnet50 --batch_size 32 --num_batches 200 --allow_growth=True

Multi-Node

Download the TensorFlow benchmark to your $WORK directory. Check out the branch that matches your tensorflow version. This example runs on two nodes in Maverick2's gtx queue (8 GPUs).

c123-456$ cdw; git clone https://github.com/tensorflow/benchmarks.git
c123-456$ git checkout cnn_tf_v2.1_compatible

Benchmark the performance with synthetic dataset on these two 2 nodes using 8 GPUs:

c123-456$ cd scripts/tf_cnn_benchmarks
c123-456$ module load intel/18.0.2 python3/3.7.0 cuda/10.1 cudnn/7.6.5 nccl/2.5.6
c123-456$ ibrun -np 8 python3 tf_cnn_benchmarks.py --variable_update=horovod \
            --num_gpus=1 --model resnet50 --batch_size 32 --num_batches 200 --allow_growth=True

TensorFlow on Frontera

These instructions detail installing and running TensorFlow benchmarks on Frontera RTX. Frontera RTX runs TensorFlow 2.1.0 with Python 3.7.0 and Intel 19. Frontera supports CUDA10.0 and CUDA/10.1. Use the appropriate CUDA version for your TensorFlow installation with Python 3.7.6.

c123-456$ module load python3 
c123-456$ module load cuda/10.1 cudnn/7.6.5 nccl/2.5.6
c123-456$ pip3 install --user grpcio==1.28.1 tensorflow-gpu==2.1.0 --no-cache-dir

We suggest installing Horovod version 0.19.2. If you wish to install other versions of Horovod, please submit a support ticket with the subject "Request for Horovod" and TACC staff will provide special instructions.

c123-456$ HOROVOD_CUDA_HOME=$TACC_CUDA_DIR HOROVOD_NCCL_HOME=$TACC_NCCL_DIR CC=gcc \
    HOROVOD_GPU_ALLREDUCE=NCCL HOROVOD_GPU_BROADCAST=NCCL HOROVOD_WITH_TENSORFLOW=1 pip3 install \
    --user horovod==0.19.2 --no-cache-dir

Single-Node

Download the tensorflow benchmark to your $WORK directory, then check out the branch that matches your tensorflow version.

c123-456$ cds; git clone https://github.com/tensorflow/benchmarks.git
c123-456$ cd benchmarks  
c123-456$ git checkout cnn_tf_v2.1_compatible

Benchmark the performance with synthetic dataset on 1 GPU

c123-456$ cd scripts/tf_cnn_benchmarks
c123-456$ module load python3/3.7.0 cuda/10.1 cudnn/7.6.5 nccl/2.5.6
c123-456$ python3 tf_cnn_benchmarks.py --num_gpus=1 --model resnet50 --batch_size 32 --num_batches 200

Benchmark the performance with synthetic dataset on 4 GPUs

c123-456$ cd scripts/tf_cnn_benchmarks
c123-456$ module load python3/3.7.0 cuda/10.1 cudnn/7.6.5 nccl/2.5.6
c123-456$ ibrun -np 4 python3 tf_cnn_benchmarks.py --variable_update=horovod --num_gpus=1 \
    --model resnet50 --batch_size 32 --num_batches 200 --allow_growth=True

Multi-Node

Download the TensorFlow benchmark to your $WORK directory. Check out the branch that matches your tensorflow version. This example runs on two nodes in the rtx queue (8 GPUs).

c123-456$ cds; git clone https://github.com/tensorflow/benchmarks.git
c123-456$ git checkout cnn_tf_v2.1_compatible

Benchmark the performance with synthetic dataset on these two 2 nodes using 8 GPUs

c123-456$ cd scripts/tf_cnn_benchmarks
c123-456$ module load python3/3.7.0 cuda/10.1 cudnn/7.6.5 nccl/2.5.6
c123-456$ ibrun -np 8 python3 tf_cnn_benchmarks.py --variable_update=horovod --num_gpus=1 \
    --model resnet50 --batch_size 32 --num_batches 200 --allow_growth=True

TensorFlow on Stampede2

These instructions detail installing and running TensorFlow benchmarks on Stampede2. Stampede2 runs TensorFlow 2.1.0 with Python 3.7 and Intel 18.

Use TACC's idev utility to grab a single compute node for 1 hour in Stampede2's skx-dev queue:

login1$ idev -p skx-dev -N 1 -n 1 -m 60

Install TensorFlow 2.1 using the default intel/18.0.2 compiler and Python 3.7:

c123-456$ module load intel/18.0.2 python3/3.7.0
c123-456$ pip3 install --user tensorflow==2.1.0 --no-cache-dir

To install horovod v0.19.2:

c123-456$ CC=gcc HOROVOD_WITH_TENSORFLOW=1 pip3 install --user horovod==0.19.2 --no-cache-dir --no-cache-dir

Single-Node

If you're not already on a compute node, then use TACC's idev utility to grab a single compute node for 1 hour:
login1$ idev -p skx-dev -N 1 -n 1 -m 60

Download the TensorFlow benchmark to your $SCRATCH directory. Check out the corresponding branch for your TensorFlow version. In this example we used cnn_tf_v2.1_compatible.

c123-456$ cd $SCRATCH
c123-456$ git clone https://github.com/tensorflow/benchmarks.git
c123-456$ cd benchmarks
c123-456$ git checkout cnn_tf_v2.1_compatible

Benchmark the performance with a synthetic dataset:

c123-456$ cd scripts/tf_cnn_benchmarks
c123-456$ export KMP_BLOCKTIME=0
c123-456$ export KMP_AFFINITY="granularity=fine,verbose,compact,1,0"
c123-456$ export OMP_NUM_THREADS=46
c123-456$ python3 tf_cnn_benchmarks.py --model resnet50 --batch_size 128 --data_format NHWC \
    --num_intra_threads 46 --num_inter_threads 2 --distortions=False --num_batches 100

Multi-Node

If you're not already on a compute node, then use TACC's idev utility to grab two compute nodes for 1 hour:

login1$ idev -p skx-dev -N 2 -n 2 -m 60

Download the TensorFlow benchmark to your $SCRATCH directory. Check out the corresponding branch for your TensorFlow version. In this example, we used cnn_tf_v2.1_compatible.

c123-456$ cd $SCRATCH
c123-456$ git clone https://github.com/tensorflow/benchmarks.git
c123-456$ git checkout cnn_tf_v2.1_compatible

Benchmark the performance with a synthetic dataset on 4 nodes:

c123-456$ module load intel/18.0.2 python3/3.7.0 
c123-456$ cd benchmarks/scripts/tf_cnn_benchmarks
c123-456$ export KMP_BLOCKTIME=0
c123-456$ export KMP_AFFINITY="granularity=fine,verbose,compact,1,0"
c123-456$ export OMP_NUM_THREADS=46
c123-456$ ibrun -np 2 python3 tf_cnn_benchmarks.py --model resnet50 --batch_size 128 \
    --variable_update horovod --data_format NCHW --num_intra_threads 46 --num_inter_threads 2 \
    --num_batches 100

The parameters for this last command are defined as follows:

  • -model specifies the neural network model
  • -batch_size specifies the number of samples in each iteration
  • -variable_update specifies using horovod to synchronize gradients
  • -data_format informs TF the nested data format comes in the order of sample count, channel, height, and width
  • -num_intra_threads specifies the number of threads used for computation within a single operation
  • -num_inter_threads specifies the number of threads used for independent operations
  • -num_batches specifies the total number of iterations to run

Tensorflow on Longhorn

Multiple versions of TensorFlow for both Python 2 and Python 3 are available through standard LMOD modules. These versions can be listed with

$ module spider tensorflow

Notice that Python 2 versions are in the tensorflow-py2 module and Python 3 versions are in the tensorflow-py3 module. Each of these modules activates the IBM PowerAI conda environment that they originate from, so your shell prompt will change when loaded.

longhorn$ module load tensorflow-py3/2.1.0
(py3_powerai_1.7.0) longhorn$

For more information about interacting with these conda environments, please refer to the Longhorn user guide.

Single-Node

Each Longhorn compute node contains four Nvidia V100 GPUs. This means one to four GPUs are available for use in a single-node job. Generic TensorFlow code can take advantage of a single GPU at a time, but multiple can be utilized either through TensorFlow's distribute module or Horovod, which comes pre-installed in each system environment.

  • Allocate a single compute node with 4 tasks (one per GPU)

    login1$ idev -N 1 -n 4 -p v100
  • Load TensorFlow 2.1.0

    c001-006$ module load tensorflow-py3/2.1.0
  • Download and checkout benchmarks compatible with TF 2.1

      (py3_powerai_1.7.0) c001-006$ git clone --branch cnn_tf_v2.1_compatible \
          https://github.com/tensorflow/benchmarks.git
      ...
      (py3_powerai_1.7.0) c001-006$ cd benchmarks
  • Run using a single GPU

      (py3_powerai_1.7.0) c001-006$ python scripts/tf_cnn_benchmarks/tf_cnn_benchmarks.py \
          --num_gpus=1 --model resnet50 --batch_size 32 --num_batches 100
      2020-06-04 16:46:56.253765: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] 
      Successfully opened dynamic library libcudart.so.10.2
      ...
      100        images/sec: 343.4 +/- 0.3 (jitter = 3.6)        7.794
      ----------------------------------------------------------------
      total images/sec: 343.19
      ----------------------------------------------------------------
      
  • Run using four GPUs using ParameterServer (no ibrun). See Distributed Training with TensorFlow.

      (py3_powerai_1.7.0) c001-006$ python scripts/tf_cnn_benchmarks/tf_cnn_benchmarks.py \
          --num_gpus=4 --model resnet50 --batch_size 32 --num_batches 100
      2020-06-04 16:53:09.078324: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] 
      Successfully opened dynamic library libcudart.so.10.2
      ...
      100        images/sec: 1269.6 +/- 0.7 (jitter = 4.4)        7.682
      ----------------------------------------------------------------
      total images/sec: 1268.75
      ----------------------------------------------------------------
      
  • Run on four GPUs using ibrun and Horovod:

      (py3_powerai_1.7.0) c001-006$ ibrun -n 4 python scripts/tf_cnn_benchmarks/tf_cnn_benchmarks.py \
          --num_gpus=1 --model resnet50 --batch_size 32 --num_batches 100 --variable_update=horovod
      TACC:  Starting up job 23600 
      TACC:  Setting up parallel environment for OpenMPI mpirun. 
      TACC:  Starting parallel tasks...
      ...
      100        images/sec: 322.6 +/- 0.4 (jitter = 3.9)        7.716
      ----------------------------------------------------------------
      total images/sec: 1289.70
      ----------------------------------------------------------------
      TACC:  Shutdown complete. Exiting.

Multi-Node

Once again, each Longhorn compute node contains four Nvidia V100 GPUs. We recommend using Horovod to scale your training or classification past a single node. This means a single process for each GPU.

  1. Allocate two compute nodes with 8 tasks (one per GPU)

     login1$ idev -N 2 -n 8 -p v100
  2. Load TensorFlow 2.1.0

     c001-006$ module load tensorflow-py3/2.1.0
  3. Download and checkout benchmarks compatible with TF 2.1

     (py3_powerai_1.7.0) c001-006$ git clone --branch cnn_tf_v2.1_compatible \
         https://github.com/tensorflow/benchmarks.git
     (py3_powerai_1.7.0) c001-006$ cd benchmarks
  4. Run using eight GPUs using ibrun and Horovod:

     (py3_powerai_1.7.0) c001-006$ ibrun -n 8 python scripts/tf_cnn_benchmarks/tf_cnn_benchmarks.py \
         --num_gpus=1 --model resnet50 --batch_size 32 --num_batches 100 --variable_update=horovod
     TACC:  Starting up job 22832
     TACC:  Setting up parallel environment for OpenMPI mpirun.
     TACC:  Starting parallel tasks...
     …
     ----------------------------------------------------------------
     total images/sec: 2560.04
     ----------------------------------------------------------------
     TACC:  Shutdown complete. Exiting.

FAQ

Q: I have missing Python packages when using TensorFlow. What shall I do?

A: Deep learning frameworks usually depend on many other packages. e.g., the Caffe package dependency list. On TACC resources, you can install these packages in user space by running:

$ pip install --user package-name