Distributed traning fail on two machine : InvalidArgumentError

Question

I have two machine, each of which has 4 GPUs. I use

with tf.device('/job:worker/replica:%d/task:%d/gpu:%d' % (FLAGS.replica_id, FLAGS.task_id, FLAGS.gpu_device_id)):

to dictate device, but failed with these error log:

tensorflow.python.framework.errors.InvalidArgumentError: Cannot assign a device to node 'init_all_tables': Could not satisfy explicit device specification '/job:worker/replica:1/task:4/device:GPU:0' because no devices matching that specification are registered in this process; available devices: /job:ps/replica:0/task:0/cpu:0, 
/job:worker/replica:0/task:0/cpu:0, /job:worker/replica:0/task:0/gpu:0, /job:worker/replica:0/task:0/gpu:1, /job:worker/replica:0/task:0/gpu:2, /job:worker/replica:0/task:0/gpu:3, /job:worker/replica:0/task:1/cpu:0, /job:worker/replica:0/task:1/gpu:0, /job:worker/replica:0/task:1/gpu:1, /job:worker/replica:0/task:1/gpu:2, /job:worker/replica:0/task:1/gpu:3, /job:worker/replica:0/task:2/cpu:0, /job:worker/replica:0/task:2/gpu:0, /job:worker/replica:0/task:2/gpu:1, /job:worker/replica:0/task:2/gpu:2, /job:worker/replica:0/task:2/gpu:3, /job:worker/replica:0/task:4/cpu:0, /job:worker/replica:0/task:4/gpu:0, /job:worker/replica:0/task:4/gpu:1, /job:worker/replica:0/task:4/gpu:2, /job:worker/replica:0/task:4/gpu:3, /job:worker/replica:0/task:5/cpu:0, /job:worker/replica:0/task:5/gpu:0, /job:worker/replica:0/task:5/gpu:1, /job:worker/replica:0/task:5/gpu:2, /job:worker/replica:0/task:5/gpu:3, /job:worker/replica:0/task:6/cpu:0, /job:worker/replica:0/task:6/gpu:0, /job:worker/replica:0/task:6/gpu:1, /job:worker/replica:0/task:6/gpu:2, /job:worker/replica:0/task:6/gpu:3, /job:worker/replica:0/task:7/cpu:0, /job:worker/replica:0/task:7/gpu:0, /job:worker/replica:0/task:7/gpu:1, /job:worker/replica:0/task:7/gpu:2, /job:worker/replica:0/task:7/gpu:3

it seems like tensorflow can't find machine B ? but I have totally same hardware and software configuration on both machine.

the start script:

# machine 10.10.12.28
~/models/inception/bazel-bin/inception/imagenet_distributed_train \
--batch_size=32 \
--data_dir=/data1/imagenet1k \
--job_name='worker' \
--replica_id=0 \
--task_id=0 \
--gpu_device_id=0 \
--ps_hosts='10.10.102.28:2220' \
--worker_hosts='10.10.102.28:2221,10.10.102.28:2222,10.10.102.28:2223,10.10.102.29:2224,10.10.102.29:2221,10.10.102.29:2222,10.10.102.29:2223,10.10.102.29:2224' &

~/models/inception/bazel-bin/inception/imagenet_distributed_train \
--batch_size=32 \
--data_dir=/data1/imagenet1k \
--job_name='worker' \
--replica_id=0 \
--task_id=1 \
--gpu_device_id=1 \
--ps_hosts='10.10.102.28:2220' \
--worker_hosts='10.10.102.28:2221,10.10.102.28:2222,10.10.102.28:2223,10.10.102.29:2224,10.10.102.29:2221,10.10.102.29:2222,10.10.102.29:2223,10.10.102.29:2224' &

~/models/inception/bazel-bin/inception/imagenet_distributed_train \
--batch_size=32 \
--data_dir=/data1/imagenet1k \
--job_name='worker' \
--replica_id=0 \
--task_id=2 \
--gpu_device_id=2 \
--ps_hosts='10.10.102.28:2220' \
--worker_hosts='10.10.102.28:2221,10.10.102.28:2222,10.10.102.28:2223,10.10.102.29:2224,10.10.102.29:2221,10.10.102.29:2222,10.10.102.29:2223,10.10.102.29:2224' &

~/models/inception/bazel-bin/inception/imagenet_distributed_train \
--batch_size=32 \
--data_dir=/data1/imagenet1k \
--job_name='worker' \
--replica_id=0 \
--task_id=3 \
--gpu_device_id=3 \
--ps_hosts='10.10.102.28:2220' \
--worker_hosts='10.10.102.28:2221,10.10.102.28:2222,10.10.102.28:2223,10.10.102.29:2224,10.10.102.29:2221,10.10.102.29:2222,10.10.102.29:2223,10.10.102.29:2224' &


CUDA_VISIBLE_DEVICES='' ~/models/inception/bazel-bin/inception/imagenet_distributed_train \
--job_name='ps' \
-task_id=0 \
--ps_hosts='10.10.102.28:2220' \
--worker_hosts='10.10.102.28:2221,10.10.102.28:2222,10.10.102.28:2223,10.10.102.29:2224,10.10.102.29:2221,10.10.102.29:2222,10.10.102.29:2223,10.10.102.29:2224' &

# machine 10.10.12.29

~/models/inception/bazel-bin/inception/imagenet_distributed_train \
--batch_size=32 \
--data_dir=/data1/imagenet1k \
--job_name='worker' \
--replica_id=1 \
--task_id=4 \
--gpu_device_id=0 \
--ps_hosts='10.10.102.28:2220' \
--worker_hosts='10.10.102.28:2221,10.10.102.28:2222,10.10.102.28:2223,10.10.102.29:2224,10.10.102.29:2221,10.10.102.29:2222,10.10.102.29:2223,10.10.102.29:2224' &

~/models/inception/bazel-bin/inception/imagenet_distributed_train \
--batch_size=32 \
--data_dir=/data1/imagenet1k \
--job_name='worker' \
--replica_id=1 \
--task_id=5 \
--gpu_device_id=1 \
--ps_hosts='10.10.102.28:2220' \
--worker_hosts='10.10.102.28:2221,10.10.102.28:2222,10.10.102.28:2223,10.10.102.29:2224,10.10.102.29:2221,10.10.102.29:2222,10.10.102.29:2223,10.10.102.29:2224' &

~/models/inception/bazel-bin/inception/imagenet_distributed_train \
--batch_size=32 \
--data_dir=/data1/imagenet1k \
--job_name='worker' \
--replica_id=1 \
--task_id=6 \
--gpu_device_id=2 \
--ps_hosts='10.10.102.28:2220' \
--worker_hosts='10.10.102.28:2221,10.10.102.28:2222,10.10.102.28:2223,10.10.102.29:2224,10.10.102.29:2221,10.10.102.29:2222,10.10.102.29:2223,10.10.102.29:2224' &

~/models/inception/bazel-bin/inception/imagenet_distributed_train \
--batch_size=32 \
--data_dir=/data1/imagenet1k \
--job_name='worker' \
--replica_id=1 \
--task_id=7 \
--gpu_device_id=3 \
--ps_hosts='10.10.102.28:2220' \
--worker_hosts='10.10.102.28:2221,10.10.102.28:2222,10.10.102.28:2223,10.10.102.29:2224,10.10.102.29:2221,10.10.102.29:2222,10.10.102.29:2223,10.10.102.29:2224' &

Answer 1

TL;DR: Don't ever use '/replica:%d' in your device specification.

The problem seems to be in your device string:

'/job:worker/replica:%d/task:%d/gpu:%d' % (FLAGS.replica_id, FLAGS.task_id, FLAGS.gpu_device_id)

The device specification '/replica:%d' is not supported in the open-source version of TensorFlow (but it is retained for some backwards compatibility reasons). The replica ID should be 0 for all tasks. You can solve this immediately by passing 0 as the --replica_id for each task, but you should really remove that flag from your version of the code.