![Horovod Example
import tensorflow as tf
import horovod.tensorflow as hvd
# Initialize Horovod
hvd.init()
# Pin GPU to be used
config = tf.ConfigProto()
config.gpu_options.visible_device_list = str(hvd.local_rank())
# Build model...
loss = ...
opt = tf.train.AdagradOptimizer(0.01)
# Add Horovod Distributed Optimizer
opt = hvd.DistributedOptimizer(opt)
# Add hook to broadcast variables from rank 0 to all other processes during initialization.
hooks = [hvd.BroadcastGlobalVariablesHook(0)]
# Make training operation
train_op = opt.minimize(loss)
# The MonitoredTrainingSession takes care of session initialization,
# restoring from a checkpoint, saving to a checkpoint, and closing when done
# or an error occurs.
with tf.train.MonitoredTrainingSession(checkpoint_dir="/tmp/train_logs",
config=config, hooks=hooks) as mon_sess:
while not mon_sess.should_stop():
# Perform synchronous training.
mon_sess.run(train_op)](https://image.slidesharecdn.com/horovod-distributedtensorflowmadeeasy-171017225750/85/Horovod-Distributed-TensorFlow-Made-Easy-12-320.jpg)
Horovod - Distributed TensorFlow Made Easy
- 1. Horovod Distributed TensorFlow Made Easy Alex Sergeev, Machine Learning Platform, Uber Engineering
- 2. Deep Learning @ Uber ● Self-Driving Vehicles ● Trip Forecasting ● Fraud Detection ● … and many more!
- 3. TensorFlow ● Most popular open source framework for Deep Learning ● Combines high performance with ability to tinker with low level model details ● Has end-to-end support from research to production
- 4. Going Distributed ● Speed up model training ● Train very large models ● Vast majority of use cases are data-parallel ● Facebook demonstrated training ResNet on ImageNet in 1 hour
- 5. Parameter Server Technique tf.Server() tf.ClusterSpec() tf.train.replicas_device_setter() tf.train.SyncReplicasOptimizer() Parameter Server Worker GPU Towers
- 6. Parameter Server Technique - Example Script Image Source: TensorFlow -- https://www.tensorflow.org/deploy/distributed
- 7. Parameter Server Technique - Performance Considering ImageNet dataset of 1.3M images, this allows to train ResNet-101 for one epoch in 3.5 minutes. Scaling efficiency on 128 GPUs is only 42%, however.
- 8. How Can We Do Better? ● Re-think necessary complexity for data-parallel case ● Improve communication algorithm ● Use RDMA-capable networking (RoCE, InfiniBand)
- 9. Meet Horovod ● Distributed training framework for TensorFlow ● Inspired by work of Baidu, Facebook, et al. ● Uses bandwidth-optimal communication protocols ○ Makes use of RDMA (RoCE, InfiniBand) if available ● Seamlessly installs on top of TensorFlow via pip install horovod ● Named after traditional Russian folk dance where participants dance in a circle with linked hands
- 10. Horovod Technique Patarasuk, P., & Yuan, X. (2009). Bandwidth optimal all-reduce algorithms for clusters of workstations. Journal of Parallel and Distributed Computing, 69(2), 117-124. doi:10.1016/j.jpdc.2008.09.002
- 11. Horovod Stack ● Plugs into TensorFlow via custom op mechanism ● Uses MPI for worker discovery and reduction coordination ● Uses NVIDIA NCCL for actual reduction on the server and across servers
- 12. Horovod Example import tensorflow as tf import horovod.tensorflow as hvd # Initialize Horovod hvd.init() # Pin GPU to be used config = tf.ConfigProto() config.gpu_options.visible_device_list = str(hvd.local_rank()) # Build model... loss = ... opt = tf.train.AdagradOptimizer(0.01) # Add Horovod Distributed Optimizer opt = hvd.DistributedOptimizer(opt) # Add hook to broadcast variables from rank 0 to all other processes during initialization. hooks = [hvd.BroadcastGlobalVariablesHook(0)] # Make training operation train_op = opt.minimize(loss) # The MonitoredTrainingSession takes care of session initialization, # restoring from a checkpoint, saving to a checkpoint, and closing when done # or an error occurs. with tf.train.MonitoredTrainingSession(checkpoint_dir="/tmp/train_logs", config=config, hooks=hooks) as mon_sess: while not mon_sess.should_stop(): # Perform synchronous training. mon_sess.run(train_op)
- 13. Horovod Example Cont. ● Run on a 4 GPU machine: ○ $ mpirun -np 4 python train.py ● Run on 4 machines with 4 GPUs each using Open MPI: ○ $ mpirun -np 16 -x LD_LIBRARY_PATH -H server1:4,server2:4,server3:4,server4:4 python train.py
- 14. Debugging - Horovod Timeline ● Discovered that ResNet-152 has a lot of tiny tensors ● Added Tensor Fusion - smart batching that gives large gains (bigger gain on less optimized networks)
- 15. Horovod Performance With Horovod, same ResNet-101 can be trained for one epoch on ImageNet in 1.5 minutes. Scaling efficiency is improved to 88%, making it twice as efficient as standard distributed TF.
- 16. Horovod Performance Cont. RDMA further helps to improve efficiency - by 30% for VGG-16.
- 17. Practical Results ● Used learning rate adjustment technique described in the Facebook paper “Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour” ● Trained convolutional networks and LSTMs in hours instead of days or weeks with the same final accuracy ● You can do that, too!
- 18. Giving Back Horovod is available on GitHub today https://github.com/uber/horovod
- 19. Thank you! Learn more about Horovod on our Eng Blog: https://eng.uber.com/horovod Learn more about ML at Uber on YouTube: http://t.uber.com/ml-meetup
- 20. Proprietary and confidential © 2017 Uber Technologies, Inc. All rights reserved. No part of this document may be reproduced or utilized in any form or by any means, electronic or mechanical, including photocopying, recording, or by any information storage or retrieval systems, without permission in writing from Uber. This document is intended only for the use of the individual or entity to whom it is addressed and contains information that is privileged, confidential or otherwise exempt from disclosure under applicable law. All recipients of this document are notified that the information contained herein includes proprietary and confidential information of Uber, and recipient may not make use of, disseminate, or in any way disclose this document or any of the enclosed information to any person other than employees of addressee to the extent necessary for consultations with authorized personnel of Uber.