IBM Watson Machine Learning CE

Getting Started

IBM Watson Machine Learning Community Edition is provided on Summit through the module ibm-wml-ce.

To access the IBM WML CE packages use the module load command:

module load ibm-wml-ce

Loading a specific version of the module is recommended to future-proof scripts against software updates. The following commands can be used to find and load specific module versions:

[user@login2.summit ~]$ module avail ibm-wml-ce

------------------------- /sw/summit/modulefiles/core --------------------------
ibm-wml-ce/1.6.1-1    ibm-wml-ce/1.6.2-0    ibm-wml-ce/1.6.2-3
ibm-wml-ce/1.6.1-2    ibm-wml-ce/1.6.2-1    ibm-wml-ce/1.7.0-1 (D)
ibm-wml-ce/1.6.1-3    ibm-wml-ce/1.6.2-2
[user@login2.summit ~]$ module load ibm-wml-ce/1.6.2-3

For more information on loading modules, including loading specific verions, see: Environment Management with Lmod

This will activate a conda environment which is pre-loaded with the following packages, and their dependencies:

IBM WML CE Version ibm-wml-ce/1.6.1 ibm-wml-ce/1.6.2 ibm-wml-ce/1.7.0
Package IBM DDL 1.4.0 IBM DDL 1.5.0 IBM DDL 1.5.1
Tensorflow 1.14 Tensorflow 1.15 Tensorflow 2.1
Pytorch 1.1.0 Pytorch 1.2.0 Pytorch 1.3.1
Caffe(IBM-enhanced) 1.0.0 Caffe (IBM-enhanced) 1.0.0 Caffe (IBM-enhanced) 1.0.0
Horovod @9f87459 (IBM-DDL Backend) Horovod v0.18.2 (IBM-DDL Backend) Horovod v0.19 (NCCL Backend)
Complete List 1.6.1 Software Packages 1.6.2 Software Packages 1.7.0 Software Packages


WML-CE was the topic of the October 2019 User Conference Call, presented by Bryant Nelson and Brad Nemanich of IBM. (slides | recording)

Running DDL Jobs

IBM DDL provides a tool called ddlrun which facilitates the launching of DDL jobs on Summit. When called inside of a bsub script, ddlrun will automatically distribute the training job across all the compute hosts in the reservation. ddlrun automatically selects default arguments optimized for performance.

Basic DDL BSUB Script

The following bsub script will run a distributed Tensorflow resnet50 training job across 2 nodes.

#BSUB -W 0:10
#BSUB -nnodes 2
#BSUB -q batch
#BSUB -J ddl_test_job
#BSUB -o /ccs/home/<user>/job%J.out
#BSUB -e /ccs/home/<user>/job%J.out

module load ibm-wml-ce

ddlrun python $CONDA_PREFIX/horovod/examples/

bsub is used to launch the script as follows:

bsub script.bash

For more information on bsub and job submission please see: Running Jobs.

Verbose mode

Using the verbose flag (-v) with ddlrun displays much more debugging information. This should be the first step to troubleshoot errors when launching a distributed job.

Setting up Custom Environments

The IBM-WML-CE conda environment is read-only. Therefore, users cannot install any additional packages that may be needed. If users need any additional conda or pip packages, they can clone the IBM-WML-CE conda environment into their home directory and then add any packages they need.


The conda environment includes a module revision number, the ‘X’ in ibm-wml-ce-1.7.0-X. The name of the active environment can be found in the prompt string, or conda env list can be used to see what conda environments are available.

$ module load ibm-wml-ce
(ibm-wml-ce-1.7.0-X) $ conda create --name cloned_env --clone ibm-wml-ce-1.7.0-X
(ibm-wml-ce-1.7.0-X) $ conda activate cloned_env
(cloned_env) $

By default this should create the cloned environment in /ccs/home/${USER}/.conda/envs/cloned_env.

To activate the new environment you should still load the module first. This will ensure that all of the conda settings remain the same.

$ module load ibm-wml-ce
(ibm-wml-ce-1.7.0-X) $ conda activate cloned_env
(cloned_env) $

Best Distributed Deep Learning Performance

Performance Profiling

There are several tools that can be used to profile the performance of a deep learning job. Below are links to several tools that are available as part of the ibm-wml-ce module.

NVIDIA Profiling Tools

The ibm-wml-ce module contains the nvprof profiling tool. It can be used to profile work that is running on GPUs. It will give information about when different CUDA kernels are being launched and how long they take to complete. For more information on using the NVIDA profiling tools on Summit, please see these slides.

Horovod Timeline

Horovod comes with a tool called Timeline which can help analyze the performance of Horovod. This is particularly useful when trying to scale a deep learning job to many nodes. The Timeline tool can help pick various options that can improve the performance of distributed deep learning jobs that are using Horovod. For more information, please see Horovod’s documentation.

PyTorch’s Autograd Profiler

PyTorch provides a builtin profiler that can be used to find bottlenecks within a training job. It is most useful for profiling the performance of a job running on a single GPU. For more information on using PyTorch’s profiler, see PyTorch’s documentation.

Reserving Whole Racks

Most users will get good performance using LSF basic job submission, and specifying the node count with -nnodes N. However, users trying to squeeze out the final few percent of performance can use the following technique.

When making node reservations for DDL jobs, it can sometimes improve performance to reserve nodes in a rack-contiguous manner.

In order to instruct BSUB to reserve nodes in the same rack, expert mode must be used (-csm y), and the user needs to explicitly specify the reservation string. For more information on Expert mode see: Easy Mode vs. Expert Mode

The following BSUB arguments and reservation string instruct bsub to reserve 2 compute nodes within the same rack:

#BSUB -csm y
#BSUB -n 85
#BSUB -R 1*{select[((LN)&&(type==any))]order[r15s:pg]span[hosts=1]cu[type=rack:pref=config]}+84*{select[((CN)&&(type==any))]order[r15s:pg]span[ptile=42]cu[type=rack:maxcus=1]}

-csm y enables ‘expert mode’.

-n 85 the total number of slots must be requested, as -nnodes is not compatible with expert mode.

We can break the reservation string down to understand each piece.

  1. The first term is needed to include a launch node in the reservation.

  2. The second term specifies how many compute slots and how many racks.

    • Here the 84 slots represents 2 compute nodes. Each compute node has 42 compute slots.
    • The maxcus=1 specifies that the nodes can come from at most 1 rack.


The following graph shows the scaling performance of the tf_cnn_benchmarks implementation of the Resnet50 model running on Summit during initial benchmark testing.


Figure 1. Performance Scaling of IBM DDL on Summit

The following LSF script can be used to reproduce the results for 144 nodes:

#BSUB -W 1:00
#BSUB -csm y
#BSUB -n 6049
#BSUB -R "1*{select[((LN) && (type == any))] order[r15s:pg] span[hosts=1] cu[type=rack:pref=config]}+6048*{select[((CN) && (type == any))] order[r15s:pg] span[ptile=42] cu[type=rack:maxcus=8]}"
#BSUB -q batch
#BSUB -o /ccs/home/user/job%J.out
#BSUB -e /ccs/home/user/job%J.out

module load ibm-wml-ce/1.6.2-2

ddlrun --nodes 18 --racks 4 --aisles 2 python $CONDA_PREFIX/tf_cnn_benchmarks/ \
    --model=resnet50 \
    --num_gpus=1 \
    --batch_size=256 \
    --num_batches=100 \
    --num_warmup_batches=10 \
    --data_name=imagenet \
    --allow_growth=True \

Troubleshooting Tips

Full command

The output from ddlrun includes the exact command used to launch the distributed job. This is useful if a user wants to see exactly what ddlrun is doing. The following is the first line of the output from the above script:

$ module load ibm-wml-ce
(ibm-wml-ce-1.6.1-1) $ ddlrun python $CONDA_PREFIX/tf_cnn_benchmarks/ --variable_update=ddl --model=resnet50
+ /autofs/nccs-svm1_sw/summit/.swci/1-compute/opt/spack/20180914/linux-rhel7-ppc64le/xl-16.1.1-3/spectrum-mpi- \
  -disable_gdr -gpu --rankfile /tmp/DDLRUN/DDLRUN.xoObgjtixZfp/RANKFILE -x "DDL_OPTIONS=-mode p:6x2x1x1 " -n 12 \
  -mca plm_rsh_num_concurrent 12 -x DDL_HOST_PORT=2200 -x "DDL_HOST_LIST=g28n14:0,2,4,6,8,10;g28n15:1,3,5,7,9,11" bash \
  -c 'source /sw/summit/ibm-wml-ce/anaconda-base/etc/profile.d/ && conda activate /sw/summit/ibm-wml-ce/anaconda-base/envs/ibm-wml-ce-1.6.1-1 \
  > /dev/null 2>&1 && python /sw/summit/ibm-wml-ce/anaconda-base/envs/ibm-wml-ce-1.6.1-1/ddl-tensorflow/examples/mnist/'

Problems Distributing Pytorch with Multiple Data Loader Workers


It is common to encounter segmenation faults or deadlocks when running distributed PyTorch scripts that make use of a DataLoader with multiple workers. A typical segfault may look something like the following:

ERROR: Unexpected segmentation fault encountered in worker.
Traceback (most recent call last):
File "/gpfs/anaconda3/envs/powerai/lib/python3.7/site-packages/torch/utils/data/", line 724, in _try_get_data
    data = self._data_queue.get(timeout=timeout)
File "/gpfs/anaconda3/envs/powerai/lib/python3.7/", line 179, in get
File "/gpfs/anaconda3/envs/powerai/lib/python3.7/", line 300, in wait
    gotit = waiter.acquire(True, timeout)
File "/gpfs/anaconda3/envs/powerai/lib/python3.7/site-packages/torch/utils/data/_utils/", line 66, in handler
RuntimeError: DataLoader worker (pid 150462) is killed by signal: Segmentation fault.

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "", line 277, in <module>
File "", line 169, in train
    for batch_idx, (data, target) in enumerate(train_loader):
File "/gpfs/anaconda3/envs/powerai/lib/python3.7/site-packages/torch/utils/data/", line 804, in __next__
    idx, data = self._get_data()
File "/gpfs/anaconda3/envs/powerai/lib/python3.7/site-packages/torch/utils/data/", line 761, in _get_data
    success, data = self._try_get_data()
File "/gpfs/anaconda3/envs/powerai/lib/python3.7/site-packages/torch/utils/data/", line 737, in _try_get_data
    raise RuntimeError('DataLoader worker (pid(s) {}) exited unexpectedly'.format(pids_str))
RuntimeError: DataLoader worker (pid(s) 150462) exited unexpectedly


The solution is to change the multiprocessing start method to forkserver (Python 3 only) or spawn. The forkserver method tends to give better performance. This Horovod PR has examples of changing scripts to use the forkserver method.

See the PyTorch documentation for more information.