tfmesos 0.0.10

Prerequisites

• For Mesos >= 1.0.0:

1. Mesos Cluster (cf: Mesos Getting Started). All nodes in the cluster should be reachable using their hostnames, and all nodes have identical /etc/passwd and /etc/group.

2. Setup Mesos Agent to enable Mesos Containerizer and Mesos Nvidia GPU Support (optional). eg: mesos-agent --containerizers=mesos --image_providers=docker --isolation=filesystem/linux,docker/runtime,cgroups/devices,gpu/nvidia

3. (optional) A Distributed Filesystem (eg: MooseFS)

4. Ensure latest TFMesos docker image (tfmesos/tfmesos) is pulled across the whole cluster

• For Mesos < 1.0.0:

1. Mesos Cluster (cf: Mesos Getting Started). All nodes in the cluster should be reachable using their hostnames, and all nodes have identical /etc/passwd and /etc/group.

2. Docker (cf: Docker Get Start Tutorial)

3. Mesos Docker Containerizer Support (cf: Mesos Docker Containerizer)

4. (optional) Nvidia-docker installation (cf: Nvidia-docker installation) and make sure nvidia-plugin is accessible from remote host (with -l 0.0.0.0:3476)

5. (optional) A Distributed Filesystem (eg: MooseFS)

6. Ensure latest TFMesos docker image (tfmesos/tfmesos) is pulled across the whole cluster

If you are using AWS G2 instance, here is a sample script to setup most of there prerequisites.

Running simple Test

After setting up the mesos and pulling the docker image on a single node (or a cluser), you should be able to use the following command to run a simple test.

$ docker run -e MESOS_MASTER=mesos-master:5050 \
    -e DOCKER_IMAGE=tfmesos/tfmesos \
    --net=host \
    -v /path-to-your-tfmesos-code/tfmesos/examples/plus.py:/tmp/plus.py \
    --rm \
    -it \
    tfmesos/tfmesos \
    python /tmp/plus.py mesos-master:5050

Successfully running the test should result in an output of 42 on the console.

Running in replica mode

This mode is called Between-graph replication in official Distributed Tensorflow Howto

Most distributed training models that Google has open sourced (such as mnist_replica and inception) are using this mode. In this mode, two kind of Jobs are defined with the names ‘ps’ and ‘worker’. ‘ps’ tasks act as ‘Parameter Server’ and ‘worker’ tasks run the actual training process.

Here we use our modified ‘mnist_replica’ as example:

1. Checkout the mnist example codes into a directory in shared filesystem, eg: /nfs/mnist

2. Assume Mesos master is mesos-master:5050

3. Now we can launch this script using following commands:

CPU:

$ docker run --rm -it -e MESOS_MASTER=mesos-master:5050 \
             --net=host \
             -v /nfs/mnist:/nfs/mnist \
             -v /etc/passwd:/etc/passwd:ro \
             -v /etc/group:/etc/group:ro \
             -u `id -u` \
             -w /nfs/mnist \
             tfmesos/tfmesos \
             tfrun -w 1 -s 1  \
             -V /nfs/mnist:/nfs/mnist \
             -- python mnist_replica.py \
             --ps_hosts {ps_hosts} --worker_hosts {worker_hosts} \
             --job_name {job_name} --worker_index {task_index}

GPU (1 GPU per worker):

$ nvidia-docker run --rm -it -e MESOS_MASTER=mesos-master:5050 \
             --net=host \
             -v /nfs/mnist:/nfs/mnist \
             -v /etc/passwd:/etc/passwd:ro \
             -v /etc/group:/etc/group:ro \
             -u `id -u` \
             -w /nfs/mnist \
             tfmesos/tfmesos \
             tfrun -w 1 -s 1 -Gw 1 -- python mnist_replica.py \
             --ps_hosts {ps_hosts} --worker_hosts {worker_hosts} \
             --job_name {job_name} --worker_index {task_index}

Note:

In this mode, tfrun is used to prepare the cluster and launch the training script on each node, and worker #0 (the chief worker) will be launched in the local container. tfrun will substitute {ps_hosts}, {worker_hosts}, {job_name}, {task_index} with corresponding values of each task.

Running in fine-grained mode

This mode is called In-graph replication in official Distributed Tensorflow Howto

In this mode, we have more control over the cluster spec. All nodes in the cluster is remote and just running a Grpc server. Each worker is driven by a local thread to run the training task.

Here we use our modified mnist as example:

1. Checkout the mnist example codes into a directory, eg: /tmp/mnist

2. Assume Mesos master is mesos-master:5050

3. Now we can launch this script using following commands:

CPU:

$ docker run --rm -it -e MESOS_MASTER=mesos-master:5050 \
             --net=host \
             -v /tmp/mnist:/tmp/mnist \
             -v /etc/passwd:/etc/passwd:ro \
             -v /etc/group:/etc/group:ro \
             -u `id -u` \
             -w /tmp/mnist \
             tfmesos/tfmesos \
             python mnist.py

GPU (1 GPU per worker):

$ nvidia-docker run --rm -it -e MESOS_MASTER=mesos-master:5050 \
             --net=host \
             -v /tmp/mnist:/tmp/mnist \
             -v /etc/passwd:/etc/passwd:ro \
             -v /etc/group:/etc/group:ro \
             -u `id -u` \
             -w /tmp/mnist \
             tfmesos/tfmesos \
             python mnist.py --worker-gpus 1