[BUG] Unable to run multi-node

[iidsample]

Describe the bug Followed the instructions provided in https://github.com/NVIDIA-Merlin/HugeCTR/tree/master/tutorial/multinode-training and setup the environment exactly as suggested. Including building HugeCTR separately with MULTI_NODE_ENABLED. However when trying to run it using run_multinode.sh receive the following error -

ORTE was unable to reliably start one or more daemons.
This usually is caused by:

* not finding the required libraries and/or binaries on
  one or more nodes. Please check your PATH and LD_LIBRARY_PATH
  settings, or configure OMPI with --enable-orterun-prefix-by-default

* lack of authority to execute on one or more specified nodes.
  Please verify your allocation and authorities.

* the inability to write startup files into /tmp (--tmpdir/orte_tmpdir_base).
  Please check with your sys admin to determine the correct location to use.

*  compilation of the orted with dynamic libraries when static are required
  (e.g., on Cray). Please check your configure cmd line and consider using
  one of the contrib/platform definitions for your system type.

* an inability to create a connection back to mpirun due to a
  lack of common network interfaces and/or no route found between
  them. Please check network connectivity (including firewalls
  and network routing requirements).

To Reproduce Steps to reproduce the behavior:

1. Build docker container using instructions provided here - https://nvidia-merlin.github.io/HugeCTR/master/hugectr_contributor_guide.html#how-to-start-your-development
2. Configured build directory in run_multinode.sh
3. bash run_multinode.sh Expected behavior Successful execution of script.

Environment (please complete the following information):

• OS: Ubuntu 18.04
• Graphic card: Nvidia P100
• CUDA version: CUDA 11.2
• Docker image - Followed the docker file provided here https://github.com/NVIDIA-Merlin/Merlin/blob/main/docker/training/dockerfile.ctr

[shijieliu]

Hi @iidsample Thanks for trying out HugeCTR! About the multinot-training tutorial, unfortunatly it's currently out of data and will be removed in next release. For now, we provide docker image in Merlin ngc which already support multi-node training for HugeCTR. You can choose to use some cluster job scheduler like srun to launch job on multinode. Thanks!

[iidsample]

Hi @shijieliu,

Thanks for your reply. Is there some way to launch without slurm. Like just on a bunch of nodes. It will be great help if you can provide some direction or steps to do so. Thank you.

[shijieliu]

The key idea for launching multi-node training in HugeCTR is to use mpi. Like https://github.com/NVIDIA-Merlin/HugeCTR/blob/master/tutorial/multinode-training/run_multinode.sh#L110 suggests. So the steps can be:

1. install and configure mpi in a bunch of nodes
2. use the docker image in Merlin ngc to lanuch container in each node. Use mpi in container to launch training.

[iidsample]

Hi @shijieliu,

Thanks for your quick reply. Unfortunately I have been having a lot trouble setting up mpi in the container to launch training. Essentially running mpirun from within the container. By any chance are you aware of a resource or have a guide about running mpi from within the container.

Thank you so much for your help.

[iidsample]

Hi,

I have been trying to run HugeCTR in distributed mode. When I try to run mpirun with dcn_2node_8gpu.py i get the following error - Runtime error: Error: the MPI total rank doesn't match the node count

I have made sure that the number of GPU's passed is correct in vvgpu parameter.

It will be great if you can help me with this.

[shijieliu]

Hi @iidsample

Could you provide more detailed log and scripts? THX!

[zehuanw]

Hi @iidsample, We are wondering if you have solved the problem? Thanks!

[kanghui0204]

Hi @iidsample ,now we have a multinode tutorial(https://github.com/NVIDIA-Merlin/HugeCTR/tree/master/tutorial/multinode-training) update, you can use script in tutorial to submit a multinode task with MPI. Please check if this update works for you.

[kanghui0204]

Hi @iidsample ， because this issue is opened for a long time ,and we will close issue now . If you have another question , you can reopen this issue , and comment.