We have seen this issue with Intel MPI as well and the solution was to set the following two variables.
export MPIR_CVAR_CH4_OFI_RANK_BITS=17
export MPIR_CVAR_CH4_OFI_TAG_BITS=$((41 - MPIR_CVAR_CH4_OFI_RANK_BITS))

The total length is 41 and can be adjusted for desired ranks and tags count.

Thanks,
Amit Ruhela
From: Zhou, Hui via discuss <discuss@mpich.org>
Sent: Thursday, April 8, 2021 10:18 AM
To: discuss@mpich.org <discuss@mpich.org>
Cc: Zhou, Hui <zhouh@anl.gov>
Subject: Re: [mpich-discuss] Supporting > 64K ranks in CH4/UCX netmod
 

Hi Min,

 

I think we can do something about it. We’ll follow-up when we have updates.

 

-- 
Hui Zhou

 

 

From: M Xie via discuss <discuss@mpich.org>
Date: Thursday, April 8, 2021 at 12:30 AM
To: discuss@mpich.org <discuss@mpich.org>
Cc: M Xie <xmxmxie@gmail.com>
Subject: [mpich-discuss] Supporting > 64K ranks in CH4/UCX netmod

Hi,

 

I am using MPICH on CH4/UCX netmod, the version is mpich-3.4.1.

 

I noticed that there is a configure parameter "--with-ch4-rank-bits" which can set the value of CH4_RANK_BITS, but seems CH4_RANK_BITS is not used in the code.

 

And I also find in the netmod/ucx/ucx_impl.h, _UCX_init_tag()/_UCX_recv_tag() use only 16 bits to set MPI rank in the ucp_tag, but this cannot differentiate correct ucp_tag when MPI ranks exceed 64K.

 

In Open MPI, 20 bits is used in pml/ucx module to set rank in ucp_tag, 20 bits for context, 24 bits for MPI tag, thus the maximum ranks in Open MPI can be 1M.

 

Is there any plan to support > 64K ranks in MPICH/CH4/UCX?

 

Thanks.

 

Min