[mvapich-discuss] mvapich support of MPI_connect and MPI_accept

[Dmitriy Lyubimov]

PS with Hydra execution (MPICH etc.) it works exactly that way: one may
spawn a single process with mpiexec that doesn't connect anywhere, pass in
the MPI port name it needs to connect to, and thus enable the system to
form larger communicators for computational grids as needed.

-d

On Mon, Jan 23, 2017 at 3:56 PM, Dmitriy Lyubimov <dlieu.7 at gmail.com> wrote:

> Thank you.
>
> Perphas a bit of a background. It is for an already developed system that
> currently uses custom resource manager and spawns mpi workers at different
> locations and times rather than using MPI CLI to log in into a given host
> and spawn a process there. Then it opens a port, connects and merges
> intracommunicator into one intercommunicator as needed. Hence, the need for
> port and connect/accept commands.
>
> I actually _don't want_ it to be logging anywhere using ssh and launch a
> process there as it assumes some possibly ssh user interactivity (of which
> there's none in the scenario i've described). I actually want it to spawn a
> single process without logging anywhere  start execution with single
> COMM_SELF=COMM_WORLD.
>
> Even that this ssh processing is suboptimal, I have tried to run what you
> said with
>
> mpirun_rsh -np=1 localhost MV2_SUPPORT_DPM=1 java -Xss4m -classpath ...
> (yes, workers themselves are java processes with a java mpi binding i am
> compiling specifically with mvapich2)
>
>  ... but was getting the same error.
>
> WARNING: Error in initializing MVAPICH2 ptmalloc library.Continuing
> without InfiniBand registration cache support.
> [cli_0]: aborting job:
> Fatal error in MPI_Init:
> Other MPI error, error stack:
> MPIR_Init_thread(514)...:
> MPID_Init(365)..........: channel initialization failed
> MPIDI_CH3_Init(356).....:
> MPIDI_CH3I_CM_Init(1748):
> MPICM_Init_UD_CM(2098)..: Couldn't allocate MR
>
> Ok maybe it cannot be run the way i want.
>
>
> On Mon, Jan 23, 2017 at 2:58 PM, Hari Subramoni <subramoni.1 at osu.edu>
> wrote:
>
>> Can you please try to run with mpirun_rsh? DPM is only supported with
>> mpirun_rsh for now.
>>
>> mpirun_rsh -np <num_procs> -hostfile <path_to_hostfile> MV2_SUPPORT_DPM=1
>> <binary>.
>>
>> Thx,
>> Hari.
>>
>> On Jan 24, 2017 4:11 AM, "Dmitriy Lyubimov" <dlieu.7 at gmail.com> wrote:
>>
>> Thank you.
>>
>> i tried to do
>>
>> mpiexec -genv MV2_SUPPORT_DPM 1
>> and cannot even get past initialization now.
>>
>> WARNING: Error in initializing MVAPICH2 ptmalloc library.Continuing
>> without InfiniBand registration cache support.
>> [cli_0]: aborting job:
>> Fatal error in MPI_Init:
>> Other MPI error, error stack:
>> MPIR_Init_thread(514)...:
>> MPID_Init(365)..........: channel initialization failed
>> MPIDI_CH3_Init(356).....:
>> MPIDI_CH3I_CM_Init(1748):
>> MPICM_Init_UD_CM(2098)..: Couldn't allocate MR
>>
>> Either i do not understand the invocation CLI or there's something else.
>> I do want to use IB support of course.
>>
>> On Sat, Jan 21, 2017 at 8:37 AM, Hari Subramoni <subramoni.1 at osu.edu>
>> wrote:
>>
>>> Hello Dmitriy,
>>>
>>> Can you please rerun after setting MV2_SUPPORT_DPM=1 on the command
>>> line?
>>>
>>> Regards,
>>> Hari.
>>>
>>> On Jan 21, 2017 4:18 AM, "Dmitriy Lyubimov" <dlieu.7 at gmail.com> wrote:
>>>
>>>> Hello,
>>>>
>>>> while trying to use MPI standard api for connecting/accepting
>>>> communicators into an intracom connection, i am getting the following
>>>> message (when trying to open an mpi port):
>>>>
>>>> [cli_0]: aborting job:
>>>> Fatal error in PMPI_Open_port:
>>>> Other MPI error, error stack:
>>>> PMPI_Open_port(123): MPI_Open_port(MPI_INFO_NULL, port=0x7f190a93a4b0)
>>>> failed
>>>> MPID_Open_port(70).: Function not implemented
>>>>
>>>> This is generally supported by mpich but it looks like not by mvapich2.
>>>> Is this right? Is there a workaround or perhaps i need to use extra build
>>>> switches to enable this functionality?
>>>>
>>>> thank you very much.
>>>> -Dmitriy
>>>>
>>>> _______________________________________________
>>>> mvapich-discuss mailing list
>>>> mvapich-discuss at cse.ohio-state.edu
>>>> http://mailman.cse.ohio-state.edu/mailman/listinfo/mvapich-discuss
>>>>
>>>>
>>
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.cse.ohio-state.edu/pipermail/mvapich-discuss/attachments/20170123/33a38770/attachment-0001.html>