[mpich-discuss] mpich2 and IPoIB

[Fabio Motezuki]

Hi,

 

We got something here, now the communications goes through infiniband. The machinefile for smpd is a bit different from the mpd.

Also tested the machinefile with 1.0.7 and works too.

 

Fabio

 

 

From: Jayesh Krishna [mailto:jayesh at mcs.anl.gov] 
Sent: Friday, October 31, 2008 2:03 PM
To: 'Fabio Kenji'
Cc: mpich-discuss at mcs.anl.gov
Subject: RE: [mpich-discuss] mpich2 and IPoIB

 

 Hi,
  Can you try specifying hostnames and interface ipaddresses in your machinefile (also see the difference in the way interface ipaddr is specified) ?

==========================================
Hostname_of_10_2_5_10 -ifhn 10.2.10.10
Hostname_of_10_2_5_11 -ifhn 10.2.10.11
==========================================

Regards,
Jayesh

-----Original Message-----
From: Fabio Kenji [mailto:kenji_japanese at ig.com.br]
Sent: Friday, October 31, 2008 6:39 AM
To: Jayesh Krishna
Cc: mpich-discuss at mcs.anl.gov
Subject: Re: [mpich-discuss] mpich2 and IPoIB

Hi,

Compiled the 1.0.8 with the following options:
./configure --enable-cxx --with-pm=smpd --prefix=<path> make && make install

to started the smpd using
<path>smpd -s

to launch the program I'm using
<path>mpiexec -n 4 -machinefile mf <program>

the mf file contains:
10.2.5.10 ifhn=10.2.10.10
10.2.5.11 ifhn=10.2.10.11

where 10.2.5.xx is ethernet and 10.2.10.xx is infiniband

and the communications are going through ethernet

Fabio

2008/10/30, Jayesh Krishna <jayesh at mcs.anl.gov>:
>
>
>
>  Hi,
>    Can you try the latest stable release (1.0.8) of MPICH2 ?
>    I just tried entering a non-existant ipaddress in the machinefile using the "-ifhn" option and my program failed (expected failure). A quick look of the smpd debug output shows that smpd is setting the right (as per the "-ifhn" args) environment before launching the process.
>    Please let us know if MPICH2 1.0.8 + smpd + "-ifhn" option does not work for you.
>
>  Regards,
>  Jayesh
>
>  -----Original Message-----
>  From: Fabio Motezuki [mailto:kenji_japanese at ig.com.br]
>  Sent: Wednesday, October 29, 2008 8:36 PM
>  To: Jayesh Krishna
>  Cc: mpich-discuss at mcs.anl.gov
>  Subject: Re: [mpich-discuss] mpich2 and IPoIB
>
>  Returning, I compiled de 1.0.8rc1 source and using mpd the ifhn in the mpd.hosts file worked well, but not with interface name resolution, just using ip directly.
>
>  In 1.0.7 I tried to use ifhn in the machinefile for smpd execution, but no effect even when using the ip address of the infiniband interface.
>
>
>  Fabio
>
>
>
>  Jayesh Krishna escreveu:
>
>  >
>  >  Hi,
>  >   Yes, you can use the "-ifhn" option in the machinefile with smpd.
>  >   Let us know if you have any problems.
>  >
>  > Regards,
>  > Jayesh
>  >
>  > -----Original Message-----
>  > From: owner-mpich-discuss at mcs.anl.gov  >
> [mailto:owner-mpich-discuss at mcs.anl.gov] On Behalf Of Rajeev Thakur  >
> Sent: Monday, October 27, 2008 2:22 PM  > To:
> mpich-discuss at mcs.anl.gov  > Subject: RE: [mpich-discuss] mpich2 and
> IPoIB  >  > Probably :-). From doing a grep in the smpd directory it
> looks like it  > might accept "-ifhn hostname"
>  >
>  > Rajeev
>  >
>  > > -----Original Message-----
>  > > From: owner-mpich-discuss at mcs.anl.gov  > >
> [mailto:owner-mpich-discuss at mcs.anl.gov] On Behalf Of Fabio Motezuki 
> > > Sent: Monday, October 27, 2008 2:12 PM  > > To:
> mpich-discuss at mcs.anl.gov  > > Subject: Re: [mpich-discuss] mpich2 and
> IPoIB  > >  > > Thanks Rajeev,  > >  > > I'll try it, perhaps is there
> any options like this for smpd?
>  > >
>  > > Fabio
>  > >
>  > >
>  > > Rajeev Thakur escreveu:
>  > > > Try specifying the interface name using ifhn= in the  > >
> mpd.hosts file as  > > > described in Sec 5.1.5 of the installation
> guide.
>  > > >
>  > > http://www.mcs.anl.gov/research/projects/mpich2/documentation/
>  > > files/mpich2-1
>  > > > .0.8-installguide.pdf
>  > > >
>  > > > Rajeev
>  > > >
>  > > >
>  > > >> -----Original Message-----
>  > > >> From: owner-mpich-discuss at mcs.anl.gov  > > >>
> [mailto:owner-mpich-discuss at mcs.anl.gov] On Behalf Of  > > Fabio
> Motezuki  > > >> Sent: Monday, October 27, 2008 5:35 AM  > > >> To:
> mpich-discuss at mcs.anl.gov  > > >> Subject: [mpich-discuss] mpich2 and
> IPoIB  > > >>  > > >> Hi all,  > > >>  > > >> I'm working on a cluster
> where each node is connected with two  > > >> networks:
>  > > >>
>  > > >> eth0 - gigabit ethernet
>  > > >> ib0 - IP over infiniband
>  > > >>
>  > > >> I would like to use the ib0 net for all mpi communications,
> but  > > >> when I start the example cpi all comunications go through
> eth0 is  > > this the  > > >> expected behavior?
>  > > >>
>  > > >> I'm launching the example program with "mpiexec  > >
> -machinefile mf -n 16  > > >> ./cpi" where the file "mf" contains the
> ip address of infiniband  > > >> cards.
>  > > >>
>  > > >> Fabio
>  > > >>
>  > > >>
>  > > >>
>  > > >
>  > > >
>  > > >
>  > >
>  > >
>  >
>

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/mpich-discuss/attachments/20081031/86059c89/attachment.htm>