Support host count in machinefile

[wavexx]

This is the initial support for an improved machinefile, as described in #7589.

• Support the syntax [user@]host[:port][*count] [bind_addr] in the
  machinefile.
• Initial code just re-writes the list fed to addprocs() by repeating the host
  as specified, so there is no user-visible change.

Includes basic documentation.

[StefanKarpinski]

This is good but I feel like having the count in front might be clearer. @JeffBezanson, any opinions?

[wavexx]

As in [count*][user@]... ?

  2*host
  2*user@host

no strong preference.

[StefanKarpinski]

Yeah, I prefer that. There's too much trailing junk in this format as it is and the number of processes seems like kind of an important number :-)

[wavexx]

Agreed. I'll revise the syntax.

[StefanKarpinski]

Is there anything preventing this PR from being merged? Can someone who uses the machinefile stuff more than I do take a look? @JeffBezanson? @amitmurthy?

[eschnett]

The usual format (used by several MPI implementations) is "hostname count", i.e. host name first.

[wavexx]

On 10/30/2014 04:08 PM, Erik Schnetter wrote:

The usual format (used by several MPI implementations) is "hostname count", i.e. host name first.

This was in the initial proposal. I'm agnostic to the actual format.

If there a machinefile spec that we can follow?

[eschnett]

See e.g. https://www.myricom.com/software/mpi-implementations-with-mx/637-how-do-i-create-a-machine-file-for-spawning-an-mpich-mx-job.html for a brief tutorial that includes an example for this format.

Open MPI is a modern MPI implementation; the relevant part of its documentation is here http://www.open-mpi.org/doc/v1.8/man1/mpirun.1.php#sect6. This extends the syntax above (which is probably still accepted for backward compatibility) to be more generic, allowing for additional information as well. The syntax is then "HOSTNAME slots=COUNT".

[wavexx]

On 10/30/2014 04:13 PM, Erik Schnetter wrote:

See e.g. https://www.myricom.com/software/mpi-implementations-with-mx/637-how-do-i-create-a-machine-file-for-spawning-an-mpich-mx-job.html for a brief tutorial that includes an example for this format.

Open MPI is a modern MPI implementation; the relevant part of its documentation is here http://www.open-mpi.org/doc/v1.8/man1/mpirun.1.php#sect6. This extends the syntax above (which is probably still accepted for backward compatibility) to be more generic, allowing for additional information as well. The syntax is then "HOSTNAME slots=COUNT".

I could work on reimplementing the parser if that's considered acceptable.

We would probably need to support user=/bind_addr= keywords, because I
don't see any explicit mention of username for login access. I also
don't see a list of standard keywords beyond slots.

This would break the existing machine file, but on the other hand at
least it would be extensible without breaking at each update.

[eschnett]

If you use ssh to access the hosts, then using ssh's syntax may be convenient: USER@HOSTNAME:PORT instead of just HOSTNAME. On the other hand, explicit keywords for user, port, bind_address would also be good. In this case I'd use the same spelling as ssh (i.e. bind_address instead of bind_addr).

[amitmurthy]

Since the hostfile is meant for accessing hosts via ssh, I think the current implementation in this PR, i.e. count*user@host:port works.

[wavexx]

AFAIK, MPICH uses the "user=" keyword when using rsh/ssh.
In this case, it would be better to stick to the same format, so that we can reuse the same hostfile.

[wavexx]

On 10/30/2014 04:30 PM, Amit Murthy wrote:

Since the hostfile is meant for accessing hosts via ssh, I think the
current implementation in this PR, i.e. count*user@host:port
works.

Having to deal with 3/4 different hostfiles for each tool
(dsh/mpi/julia) is a bit annoying.

I definitely support the idea of using an existing file format for this.
MPICH/OpenMPI formats seem reasonable and widely used.

[amitmurthy]

Stefan / @JeffBezanson should take a decision and close this out. My preference is still for count*user@host:port, purely from a compactness point of view.

[StefanKarpinski]

My preference is still for count*user@host:port, purely from a compactness point of view.

Me too — we can have a translation or compatibility mode for other formats, which are IMO less nice – and since there's more than one of them there's no way to make everyone happy.

[StefanKarpinski]

I'm doing some work on a branch and I'd rather not futz with it – would you mind resolving the conflict here and merging this?

[amitmurthy]

OK. Will do.

[StefanKarpinski]

Thank you :-)

[StefanKarpinski]

Thanks @amitmurthy and @wavexx – sorry that took so long to merge.