[Beowulf] Does anyone here mix CISC and RISC within their clusters.

James Cownie jcownie at cantab.net
Fri Oct 28 11:04:14 PDT 2016


> no idea about how MPI should behave in this setup
Provided you describe your messages correctly (e.g. as 1000 32 bit ints, rather than as 4000 bytes), then MPI shoul do the right thing.
(Including cases like moving floats from a little-endian IEEE machine [X86] to a big-endian IBM format machine [IBM 360 et seq]).

Of course, the “provided” is the big issue there. If your code pretends just to send bytes when the data is more complicated then all bets are off.
--
-- Jim
--
James Cownie <jcownie at cantab.net>

> On 27 Oct 2016, at 06:30, C Bergström <cbergstrom at pathscale.com> wrote:
> 
> Could this be related some some ENDIAN issue? I've never heard of this
> before and quite frankly - no idea about how MPI should behave in this
> setup.. curious..
> 
> What's the exact error and have you tried asking the openmpi or mpich
> mailing lists?
> 
> On Thu, Oct 27, 2016 at 1:00 PM, Darren Wise <darren at wisecorp.co.uk> wrote:
>> Heya folks,
>> 
>> I have a SUN T2000 SPARC64 quadcore (16 HW threads niagra) using ORACLE
>> Linux 1.0 (don't ask but SOLARIS 10,11 just was not having it.. At all)
>> 
>> Along with seven dual socket, quadcore AMD x86-64 CISC nodes running ubuntu
>> 16.4LTS, MPICH and OpenMPI are giving me some strange errors but as soon as
>> I opt out the SUN box everything runs smoothly.
>> 
>> I know it's not the best idea since sliced bread came about, but if I have
>> the machines around just wanted to add the sun T2k into the mix.
>> 
>> Tried starting processes from both CISC and RISC sides but only when they
>> are seperate do they fully complete.
>> 
>> Shared NFS and passwordless SSH is all installed fine apart from some
>> version number differences and the same with the actual age of the
>> distributions, ORACLE linux SPARC64 edition version 1.0 must be a few years
>> old now circa 2013 maybe verses ubuntu 16.4LTS which must be early on this
>> year and the same with OpenMPI versions differing I gather..
>> 
>> I will try to down and upgrade the packages from source, compile and try to
>> sync all that.
>> Just curious really if anyone else has tried doing this at home or where
>> ever..
>> 
>> 
>>> Kind regards,
>>> Darren Wise Esq,
>>> B.Sc, HND, GNVQ, City & Guilds.
>>> 
>>> Managing Director (MD)
>>> Art Director (AD)
>>> Chief Architect/Analyst (CA/A)
>>> Chief Technical Officer (CTO)
>>> 
>>> www.wisecorp.co.uk
>>> www.wisecorp.co.uk/babywise
>>> www.darrenwise.co.uk
>> 
>> _______________________________________________
>> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
>> To change your subscription (digest mode or unsubscribe) visit
>> http://www.beowulf.org/mailman/listinfo/beowulf
>> 
> _______________________________________________
> Beowulf mailing list, Beowulf at beowulf.org sponsored by Penguin Computing
> To change your subscription (digest mode or unsubscribe) visit http://www.beowulf.org/mailman/listinfo/beowulf





More information about the Beowulf mailing list