[mvapich-discuss] Configuration of version 2.3 with TCP fallback?

Petr Hanousek petr.hanousek at cesnet.cz
Wed Jun 7 08:14:42 EDT 2017


Hello Hari,

On 6.6.2017 18:30, Hari Subramoni wrote:
> Hi,
> 
> At this point, the default MVAPICH channel cannot use TCP as a
> fallback if IB fails.

OK, then I solve my issue either with deprecated Nemesis or by compiling
two versions of MVAPICH. In both cases there should be some logic before
using the module because of an "error and exit" behavior of Nemesis when
IB fails.

> Please note that IB is a pretty resilient fabric and it is unlikely to
> fail. Thus you may never need such a fallback option.

I agree. But I need to use MVAPICH in quite heterogenous grid
environment. Many different clusters interconnected by a common
scheduler (PBS Pro). Not every cluster has IB cards or is connected to
others by them. Therefore I need to have some fallback possibility.

Petr

> 
> On Tue, Jun 6, 2017 at 7:02 AM, Petr Hanousek <petr.hanousek at cesnet.cz
> <mailto:petr.hanousek at cesnet.cz>> wrote:
> 
>     Dear all,
>     in my previous mail I was really "out" when I tried to use
> 
>     ./configure --prefix=/software/mvapich/2.3/gcc --enable-cxx
>     --enable-fortran=all --with-pm=hydra --with-device=ch3:mrail:ib,tcp
>     --enable-pbs-launcher --with-ibverbs=/software/ofed-1.5.4
>     --enable-shared
> 
>     After careful read of documentation and some googling around and some
>     test I assume my --with-device parameter is bad. And it was caused by my
>     misunderstanding of the whole concept. So my question now is if mvapich
>     2.3 can be configured to use IB with TCP fallback if IB fails, when
>     Nemesis is deprecated?
> 
>     Cheers Petr

-- 
+-------------------------------------------------------------------+
   Petr Hanousek                   e-mail: petr.hanousek at cesnet.cz
   User Support                    phone:  +420 950 072 112
   CESNET z.s.p.o.                 mobile: +420 606 665 139
   location: Zikova 4, Praha       room: 90a
                        Czech Republic
+-------------------------------------------------------------------+

-------------- next part --------------
A non-text attachment was scrubbed...
Name: smime.p7s
Type: application/pkcs7-signature
Size: 3718 bytes
Desc: S/MIME Cryptographic Signature
URL: <http://mailman.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20170607/a839237c/attachment-0001.p7s>


More information about the mvapich-discuss mailing list