[Pkg-openmpi-maintainers] Bug#845594: openmpi-bin: MPI_Comm_spawn fails if blt_tcp_if_include is set to lo

[Thibaut Paumard]

Dear Alastair,

I confirm the package in experimental (2.0.2~git.20161225-2) fixes this bug.

For the record, this new version forced me to use a hostfile. It looks
like the old default was to provide a large number of slots on
localhost, while the new default is no slots at all. Not a big deal though.

Regards, Thibaut.

Le 02/12/2016 à 16:52, Alastair McKinstry a écrit :
> Hi,
> 
> How have you been turning "networking on/off" ?
> 
> I have been testing this using a local version of head-of-tree (soon to
> be openmpi 2.0.2). This is head-of-tree as of Nov 25, with some local
> fixes needed for Debian - adding libpmix* which was not fully included;
> For me, the test case fails as advertised for unstable, works for me (tm).
> 
> I'm uploading my working version to experimental for further tests.
> 
> regards
> Alastair
> 
> 
> On 01/12/2016 15:04, Thibaut Paumard wrote:
>> Hi,
>>
>> I have made a couple of additional tests.
>>
>> I have run
>>  mpirun --mca btl_tcp_if_include lo --mca btl tcp,self -np 2 \
>>         --mca plm_rsh_agent /bin/true slave
>> under a variety of environments. I get a surprising result: it works in
>> a sid chroot in some circumstances.
>>
>> Results:
>>
>> - plain jessie, networking on:                   PASS
>> - plain jessie, networking off:                  PASS
>> - jessie VirtualBox, networking on:              PASS
>> - jessie VirtualBox, networking off:             PASS
>> - jessie chroot (jessie host), networking on:    PASS
>> - jessie chroot (jessie host), networking off:   PASS
>>
>> - unstable chroot (jessie host), networking on:  PASS
>> - unstable chroot (jessie host), networking off: FAIL
>> - unstable VirtualBox:                           FAIL
>>
>> I attach the error log from the attempt "unstable chroot (jessie host),
>> networking off".
>>
>>
>> Kind regards, Thibaut.
>>
>> Le 25/11/2016 à 08:51, Thibaut Paumard a écrit :
>>> Control: retitle 845594 "openmpi: lo interface broken in the tcp btl"
>>>
>>> Hi,
>>>
>>> Actually the regression can also be demonstrated without using
>>> MPI_Comm_spawn with:
>>>  mpirun -np 2 --mca btl_tcp_if_include lo --mca btl tcp,self ./slave
>>>
>>> The above command runs fine under jessie (openmpi 1.6.5-9.1) but fails
>>> under sid.
>>>
>>> For the record my test environment is a production machine for jessie
>>> and a virtualbox virtual machine for sid.
>>>
>>> Kind regards, Thibaut.
>>>
>