Do not mark a -host specification as "slots given" unless it actually specifies the #slots

[rhc54]

This allows the system to dynamically detect the number of available processors and set the #slots accordingly.

Signed-off-by: Ralph Castain rhc@open-mpi.org

[hppritcha]

Do we need to change any documentation for this change in behavior?

[rhc54]

Good point - I have updated the man page

[hppritcha]

@jsquyres what do you think about this kind of change in behavior within a release stream?
I recall a lot of discussion about behavior of --host vs --hostfile etc.
at one of our recent F2F's:
https://github.com/open-mpi/ompi/wiki/Meeting-2016-02
just wanting to make sure we are keeping with the original decisions.

[hppritcha]

Here's the spread sheet that we worked on at the Dallas F2F 2/16:

https://docs.google.com/spreadsheets/d/1poOwNKtYxnDnpF7-D15lmcFVtLRNrmf4_hu3Obtu95M/edit

[rhc54]

@hppritcha Hate to tell you, but that spreadsheet isn't accurate - at least, that isn't the current behavior.

[ggouaillardet]

@hppritcha @rhc54 imho, this specific case is kind of undocumented/unspecified.

one one hand, we specify a slot-list, and on the other hand, we specify a host.

for example, what if the slot-list contains 12 slots but we specify a different number of slots with the --host option ? should we try to handle this in a "nice" way ? or should we simply abort because of an inconsistency in the options requested by the user ?

in this very specific case, e.g. mpirun --slot-list 0:0-5,1:0-5 --host xxx, yet an other option could be to set the ORTE_NODE_FLAG_SLOTS_GIVEN flag, and set the number of slots to 12 (e.g. the number of slots in the list)

[rhc54]

I've said this before, but I'll reiterate here - I am no longer supporting the -host/-hostfile code. We keep coming up with every imaginable corner-case, and the complexity of trying to handle all of them - while preserving existing behavior - is insane. Clear evidence: the current master behavior no longer mirrors @jsquyres spreadsheet.

This patch makes the branch follow the master and fixes the user's problem, but the result isn't the behavior in the spreadsheet. Someone else can figure out what they want to do.

[jladd-mlnx]

bot:mellanox:retest

[jladd-mlnx]

bot:mellanox:retest

[jladd-mlnx]

bot:mellanox:retest

[hppritcha]

Here's what we have at the moment across several releases for the --host/---hostfile behavior where host_foo has a single entry foo. The table below shows how many processes are launched as a function of mpirun command line options and Open MPI release:

Release	mpirun --host foo	mpirun --host foo:2	mpirun --hostfile host_foo	mpirun --hostfile host_foo -np 1
1.8.8	one process	doesn't work	nslot processes	one process
1.10.0	one process	doesn't work	nslot processes	one process
1.10.5	one process	doesn't work	nslot processes	one process
1.10.x	one process	doesn't work	nslot processes	one process
2.0.0	one process	two processes	nslot processes	one process
2.0.1	one process	two processes	nslot processes	one process
3.0.0a1	nslot processes	two processes	nslot processes	one process

where nslots is the number of cores on the host foo

[hppritcha]

Discussed at the devel F2F and decided to keep the behavior of 1.10.x series (modulo the new --host foo:X). So closing this PR.

[jsquyres]

For reference: https://github.com/open-mpi/ompi/wiki/Meeting-2017-01

[jjhursey]

Food for thought: It would be nice if we had a system regularly running a handful of -host and -hostfile runs and compare the output to confirm that we stay in compliance with what we intend to do. Maybe a CI test on a specific system (keeping to a specific system makes it easy to setup without handling the generic system problem).