Well, it turns out that I won't be able to do this after all. :-/
The problem is that the new hostfile handling system gets a tad upset
if it thinks a hostfile has been specified, but then it doesn't find
anything in it. I started trying to patch it to allow for the default
hostfile name to be a special case, but it winds up requiring patches
in a number of places - gets really ugly.
So I'm afraid the bottom line is that someone is going to have to
enter the name of the default hostfile if they really have one they
want us to use. Or put it in their environment.
I imagine you could do something in your code to check and set the MCA
param yourself if you want. Not sure if it's more trouble than it's
worth - have to leave that one to you.
Sorry...
Ralph
On Dec 15, 2008, at 10:05 AM, Ralph Castain wrote:
<laugh> can't argue with that logic! I'll make the change and file
the CMR.
Thanks
Ralph
On Dec 15, 2008, at 9:06 AM, Greg Watson wrote:
Hi Ralph,
I think mainly because it simplifies installation of ompi for PTP
users. Since PTP uses the hostfile to display the system
configuration, we're pretty much always going to have one (although
PTP does work without it, feedback is more limited). It's much
easier for people to add a list of hosts to a file, than have to go
and add something to the param file as well (it's hard enough to
get them to do the former correctly).
Greg
On Dec 15, 2008, at 1:51 PM, Ralph Castain wrote:
Can you help me understand something here? I'm not opposed to
making the change - just puzzled as to why the value of the
default hostfile name is of concern to Eclipse.
There is one reason not to set a default name - it causes us to
open and read that file, even though nobody ever put something in
it. Remember, we distribute and install an empty default hostfile
that contains instructions on how to build one, so it will always
exist. Since the name of the default hostfile can be set in the
default MCA param file, environment, or cmd line, there didn't
seem to be any real reason to define some special name.
It isn't a big deal, though, so I don't really care that much. But
I would like to understand why Eclipse cares so we can factor that
into any future thinking.
Ralph
On Dec 12, 2008, at 7:11 AM, Greg Watson wrote:
From our perspective, it would be good if it could default to the
old behavior (in 1.3 if possible).
Thanks,
Greg
On Dec 8, 2008, at 11:42 AM, Ralph Castain wrote:
I don't think there was any overt thought given to it, at least
not on my part. I suspect it came about because (a) the wiki
defining hostfile behavior made no mention of the default value,
(b) I may have overlooked the prior default when rewriting that
code, and (c) since we now have default-hostfile as well as
hostfile, it could be I didn't default the name since it isn't
clear which one should get the default.
I honestly don't remember - this has been in the code base for a
really long time now.
I have no iron in this fire - as you know, all of our environs
here are managed. So I guess I'll throw it out there to the
community:
do we want --default-hostfile to have a default value?
Pros: it could be considered a continuation of 1.2's hostfile
behavior
Cons: we treat hostfile in a totally different way in 1.3. We
now have two hostfiles: a default that applies to all
app_contexts, and a hostfile that applies to only one
app_context. It would seem that the default-hostfile best aligns
with the old "hostfile" behavior, but could lead to some
confusion in its new usage.
Any preferences/thoughts?
Ralph
On Dec 5, 2008, at 9:15 AM, Greg Watson wrote:
Hi,
In 1.2.x, the rds_hostfile_path parameter pointed to openmpi-
default-hostfile by default. This parameter has been replaced
with orte_default_hostfile in 1.3, but now it defaults to
<none>. Was there any particular reason for the default value
to change?
Greg
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
_______________________________________________
devel mailing list
de...@open-mpi.org
http://www.open-mpi.org/mailman/listinfo.cgi/devel
