The script doesn't create a correct Standard_HC44rs cluster with Mellanox EDR InfiniBand about azurehpc

[jithinjosepkl]

AN (Accelerated networking) is not available for HC. Please set "accelerated_networking": false for HC

[Smahane]

@jithinjosepkl in the config file i shared? Will that fix the interconnect issue?

[jithinjosepkl]

Please follow this article to make IMPI pick the mlx provider.

Next update of CentOS HPC images will include IMPI 2019-U8, where you don't have to specify this environment parameter (picks up mlx provider by default).

[Smahane]

@jithinjosepkl I would appreciate a working scenario as the cost is getting very high without us getting a job ran correctly.
I'm interested in Standard_HC44rs cluster with Mellanox and Gen2 VM

surprisingly the OpenLogic:CentOS:7_7-gen2:latest image didn't come with any IMPI installed (It sounds like it get overridden maybe via the azurehpc scripts?). So i had to install IMPI manually .

When i setup the Mlx, i get an error because the interconnect is not installed/setup correctly:

[hpcadmin@headnode lammps-avx512]$ export FI_PROVIDER=mlx
[hpcadmin@headnode lammps-avx512]$ mpirun -n 2 -ppn 44 /opt/intel/psxe_runtime/linux/mpi/intel64/bin/IMB-MPI1 pingpong
[0] MPI startup(): Intel(R) MPI Library, Version 2019 Update 8  Build 20200624 (id: 4f16ad915)
[0] MPI startup(): Copyright (C) 2003-2020 Intel Corporation.  All rights reserved.
[0] MPI startup(): library kind: release
[0] MPI startup(): libfabric version: 1.10.1-impi
Abort(1091215) on node 0 (rank 0 in comm 0): Fatal error in PMPI_Init_thread: Other MPI error, error stack:
MPIR_Init_thread(136)........:
MPID_Init(1138)..............:
MPIDI_OFI_mpi_init_hook(1061): OFI addrinfo() failed (ofi_init.c:1061:MPIDI_OFI_mpi_init_hook:No data available)

[garvct]

The CentOS images do not contain any pre-installed OFED drivers and mpi libraries (including Intel MPI), try using the CentOS-HPC images instead (e.g OpenLogic:CentOS-HPC:7_7-gen2:latest). The CentOS-HPC images should be ready to go if you want to use Infiniband on HC44 skus.

[Smahane]

@garvct I'm already using OpenLogic:CentOS-HPC:7_7-gen2:latest image. Please see my original post

[jithinjosepkl]

@Smahane , based on your config file, you are using OpenLogic:CentOS:7_7-gen2:latest.
"hpc_image": "OpenLogic:CentOS:7_7-gen2:latest",

You need OpenLogic:CentOS-HPC:7_7-gen2:latest image instead for the MPIs to be pre-installed.

[garvct]

Once you sort out your image, you can execute the IMB-MPI1 benchmark using the scripts in azurehpc/apps/imb-mpi. Examples of running IMB-MPI1 with different MPI libraries are provided.

[Smahane]

@garvct and @jithinjosepkl thank you for pointing me out of this HPC image.

• Should headnode have this image too?
• Should I still set "accelerated_networking": false" ? And what does it mean?
• any instructions of how to add my own post installation script to the config file?

[xpillons]

@Smahane

• The Headnode can have the HPC image too
• Accelerated Networking is offloading TCP and will boost the frontend NIC, but is not available on all VM SKUs. See the public documentation here. Accelerated Networking doesn't apply to the Infiniband NIC and today is not supported on our HPC VM SKUs yet.
• You can add your own scripts by
  • Adding your scripts in the scripts directory where your config file is stored
  • Add a custom tag on the resources you want your scripts to be applied
  • In the install array add a section for each script you want to be applied, specify if sudo is required, and add params and deps files if needed

[Smahane]

This worked. Thank you everyone

[Smahane]

Hello @xpillons @garvct and @jithinjosepkl . I ran LAMMPS on up to 8 nodes Standard_HC44rs but I'm having performance issues at 8 nodes:

I think one or more nodes are bad. Do you know of any smoke test or script that can help me check the nodes and detect which one is not working well?

the azhpc_install_config/install/11_node_healthchecks.log doesn't show any errors.

Thank you,

[xpillons]

@Smahane you can use the MPI PingPong test, we have an example here https://github.com/Azure/azurehpc/tree/master/apps/imb-mpi