Hi Geoff, This is kind of what I suspected. I think its a very bad design decision to have the open mpi runtime under the hood setting Open MP environment variables. At the very minimum, there should be an mca parameter to over ride this, or alternatively, this section of code would only be compiled based on compiler specific sentinel macros.
The whole open MP env. variable thing is kind of a mess. I know this based on experience at my former employer working with a variety of open MP compiler runtimes to make sure they worked well with my former employer's alps runtime system. This should have been an RFC, not something coming in on a Sunday without a review request. I was using PGI 12.9 on carver. What I absolutely do not like is having something like this coming in on a weekend without a review then having ex post facto to spend time cleaning up. Anyway, a wall will be put in place around the v2.x release to avoid these kind of surprises there. Needless to say I will not be admitting this PR in to v2.x until its cleaned up enough to work with all major compilers, or else is only activated when OMPI is compiled with an Intel compiler. Howard 2015-06-30 16:00 GMT-06:00 Geoffrey Paulsen <gpaul...@us.ibm.com>: > I discussed with Robert Ho who was working with Ralph on this option. He > believes it's possible that the PGI compiler / runtime does not understand > OMP_PROC_BIND=SPREAD which was only introduced in OpenMP 4.0. > > Unfortunately I can't find any docs as the http://www.pgroup.com/index.htm is > down right now. > > We have PGI version 11.8 which only support OpenMP version 3.0, and does > not list OMP_PROC_BIND at all. > > in 11.8, PGI supported MP_BIND=yes which would request the PGI runtime > libraries to bind processes or threads in a parallel region to phsyical > processors (default is no). > It also supported MP_BLIST=a,b,c,d (when MP_BIND was set to yes to map > how you wanted threads or processes bound to physical processors 0,1,2,3. > > There is a note in the documentation that setting MP_BIND does NOT affect > the compiler behavior at all, only the runtime library. > > > Regards, > > *Geoffrey (Geoff) Paulsen* > Software Engineer - Platform MPI > ------------------------------ > *Phone:* 1-720-349-2832 > * E-mail:* *gpaul...@us.ibm.com* <gpaul...@us.ibm.com> > [image: IBM] > > 1177 S Belt Line Rd > Coppell, TX 75019-4642 > United States > > > [image: Inactive hide details for Howard Pritchard ---06/29/2015 09:27:12 > PM---I decided just to disable the carver/pgi mtt runs. 2015-]Howard > Pritchard ---06/29/2015 09:27:12 PM---I decided just to disable the > carver/pgi mtt runs. 2015-06-29 15:10 GMT-06:00 Ralph Castain <rhc@ope > > From: Howard Pritchard <hpprit...@gmail.com> > To: Open MPI Developers <de...@open-mpi.org> > Date: 06/29/2015 09:27 PM > Subject: Re: [OMPI devel] OMPI_PROC_BIND value is invalid errors > Sent by: "devel" <devel-boun...@open-mpi.org> > ------------------------------ > > > > I decided just to disable the carver/pgi mtt runs. > > > 2015-06-29 15:10 GMT-06:00 Ralph Castain <*r...@open-mpi.org* > <r...@open-mpi.org>>: > > Very strange then - again, can you run it with the verbose flag and > send me the output? I can't replicate what you are seeing. > > > On Mon, Jun 29, 2015 at 4:05 PM, Howard Pritchard < > *hpprit...@gmail.com* <hpprit...@gmail.com>> wrote: > ibm dataplex and laki ~= cray. nothing to do with cray. > Cray runs fine since I use aprun there. > > > 2015-06-29 13:54 GMT-06:00 Ralph Castain <*r...@open-mpi.org* > <r...@open-mpi.org>>: > Hmmm...is this some Cray weirdness? I checked the code and it > looks right, and it runs correctly for me on both Mac and Linux. All > it is > doing is calling "setenv", so I'm wondering if there is something > environ-specific going on here? > > I added some debug in cast that might help - can you run it on > the Cray with "--mca rtc_base_verbose 5" on the cmd line? > > > On Mon, Jun 29, 2015 at 1:19 PM, Jeff Squyres (jsquyres) < > *jsquy...@cisco.com* <jsquy...@cisco.com>> wrote: > Ahh... it's OMP_PROC_BIND, not OMPI_PROC_BIND. > > Yes, Ralph just added this. > > I chatted with him about this on the phone moments ago; he's > pretty sure he knows where to go look to find the problem. > > > > On Jun 29, 2015, at 12:00 PM, Howard Pritchard < > *hpprit...@gmail.com* <hpprit...@gmail.com>> wrote: > > > > laki is also showing the errors: > > > > > > Here's the shortened url: > > > > *http://goo.gl/Ra264U* <http://goo.gl/Ra264U> > > > > looks like the badness started with the latest nightly. > > I think there was some activity in the orte binding area > recently. > > > > Howard > > > > > > > > > > 2015-06-29 9:52 GMT-06:00 Jeff Squyres (jsquyres) < > *jsquy...@cisco.com* <jsquy...@cisco.com>>: > > Can you provide an MTT short URL to show the results? > > > > Or, if the MTT results are not on the community reporter, > can you show a bit more context in the output? > > > > > > > On Jun 29, 2015, at 11:47 AM, Howard Pritchard < > *hpprit...@gmail.com* <hpprit...@gmail.com>> wrote: > > > > > > Hi Folks, > > > > > > I'm seeing an error I've not seen before in the MTT runs > on the ibm dataplex > > > at NERSC. The mpirun launched jobs are failing with > > > > > > OMPI_PROC_BIND value is invalid > > > > > > errors. > > > > > > This is is for the trivial ring tests. > > > > > > Is anyone else seeing these types of errors? > > > > > > Howard > > > > > > _______________________________________________ > > > devel mailing list > > > *de...@open-mpi.org* <de...@open-mpi.org> > > > Subscription: > *http://www.open-mpi.org/mailman/listinfo.cgi/devel* > <http://www.open-mpi.org/mailman/listinfo.cgi/devel> > > > Link to this post: > *http://www.open-mpi.org/community/lists/devel/2015/06/17558.php* > <http://www.open-mpi.org/community/lists/devel/2015/06/17558.php> > > > > > > -- > > Jeff Squyres > > *jsquy...@cisco.com* <jsquy...@cisco.com> > > For corporate legal information go to: > *http://www.cisco.com/web/about/doing_business/legal/cri/* > <http://www.cisco.com/web/about/doing_business/legal/cri/> > > > > _______________________________________________ > > devel mailing list > > *de...@open-mpi.org* <de...@open-mpi.org> > > Subscription: > *http://www.open-mpi.org/mailman/listinfo.cgi/devel* > <http://www.open-mpi.org/mailman/listinfo.cgi/devel> > > Link to this post: > *http://www.open-mpi.org/community/lists/devel/2015/06/17559.php* > <http://www.open-mpi.org/community/lists/devel/2015/06/17559.php> > > > > _______________________________________________ > > devel mailing list > > *de...@open-mpi.org* <de...@open-mpi.org> > > Subscription: > *http://www.open-mpi.org/mailman/listinfo.cgi/devel* > <http://www.open-mpi.org/mailman/listinfo.cgi/devel> > > Link to this post: > *http://www.open-mpi.org/community/lists/devel/2015/06/17560.php* > <http://www.open-mpi.org/community/lists/devel/2015/06/17560.php> > > > -- > Jeff Squyres > *jsquy...@cisco.com* <jsquy...@cisco.com> > For corporate legal information go to: > *http://www.cisco.com/web/about/doing_business/legal/cri/* > <http://www.cisco.com/web/about/doing_business/legal/cri/> > > _______________________________________________ > devel mailing list > *de...@open-mpi.org* <de...@open-mpi.org> > Subscription: > *http://www.open-mpi.org/mailman/listinfo.cgi/devel* > <http://www.open-mpi.org/mailman/listinfo.cgi/devel> > Link to this post: > *http://www.open-mpi.org/community/lists/devel/2015/06/17563.php* > <http://www.open-mpi.org/community/lists/devel/2015/06/17563.php> > > > _______________________________________________ > devel mailing list > *de...@open-mpi.org* <de...@open-mpi.org> > Subscription: > *http://www.open-mpi.org/mailman/listinfo.cgi/devel* > <http://www.open-mpi.org/mailman/listinfo.cgi/devel> > Link to this post: > *http://www.open-mpi.org/community/lists/devel/2015/06/17567.php* > <http://www.open-mpi.org/community/lists/devel/2015/06/17567.php> > > > _______________________________________________ > devel mailing list > *de...@open-mpi.org* <de...@open-mpi.org> > Subscription: *http://www.open-mpi.org/mailman/listinfo.cgi/devel* > <http://www.open-mpi.org/mailman/listinfo.cgi/devel> > Link to this post: > *http://www.open-mpi.org/community/lists/devel/2015/06/17568.php* > <http://www.open-mpi.org/community/lists/devel/2015/06/17568.php> > > > _______________________________________________ > devel mailing list > *de...@open-mpi.org* <de...@open-mpi.org> > Subscription: *http://www.open-mpi.org/mailman/listinfo.cgi/devel* > <http://www.open-mpi.org/mailman/listinfo.cgi/devel> > Link to this post: > *http://www.open-mpi.org/community/lists/devel/2015/06/17569.php* > <http://www.open-mpi.org/community/lists/devel/2015/06/17569.php> > > _______________________________________________ > devel mailing list > de...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2015/06/17570.php > > > _______________________________________________ > devel mailing list > de...@open-mpi.org > Subscription: http://www.open-mpi.org/mailman/listinfo.cgi/devel > Link to this post: > http://www.open-mpi.org/community/lists/devel/2015/06/17571.php >
