Re: [OMPI users] Why should the attached code wait on MPI_Bcast

[s anwar]

Thanks.

On 7/18/06, Bert Wesarg  wrote:


hi,

s anwar wrote:
> Thank you for the clarification. Why is MPI_COMM_SELF not the correct
> communicator for MPI_Comm_spawn(). My application will have a single
> master only.
yes for a single master this should be the same, but never try this.
>
> Also, can I merge the intercommunicator into an intracommunicator via
> MPI_Intercomm_merge(intercomm, 0, ) and use MPI_Bcast(..., 0,
> intracomm) instead of sending MPI_ROOT or 0 (root's rank in local
group).
dito, never try this, but it should work

bye
bert wesarg
>
> Thanks.
>
> On 7/18/06, *Bert Wesarg*  > wrote:
>
> hi,
>
> yes sorry for my first reply, my words were to rough.
>
> a bcast for a intercomm works this way (in your words):
>
> - your masters want to send a buffer to your slaves
> - one of the masters must provide the MPI_ROOT as root in the
> MPI_BCAST call
> - all slaves must provide the rank of this MPI_ROOT as the root
argument
> - all others ranks from the masters, must provide MPI_PROC_NULL as
root
>
> so the buffer will be send from MPI_ROOT to all slaves
>
> if this isn't that clear, you should understand the concept of mpi
> intercomms. a good starting point could be the standard:
> http://www.mpi-forum.org/docs/mpi-11-html/node111.html
> 
>
> by the way:
> * your one_int is never intitialized
> * i don't know if the MPI_COMM_SELF is the right comm for the
> MPI_COMM_SPAWN
>
> by
> bert wesarg
>
> s anwar wrote:
> > I don't think I blamed the implementation in any way in my
original
> > email. My intent is to gain understanding of why my code
> does/should not
> > work. I assumed that I was not passing the correct
intercommunicator.
> > However, I am at a loss on how to construct a proper
> intercommunicator
> > in this case. You have mentioned that I haven't defined any group
> to be
> > the root group. Could you care to elaborate on how can I make a
> group a
> > root group?
> >
> > Thanks.
> >
> > On 7/18/06, *Bert Wesarg* < wes...@informatik.uni-halle.de
> 
> >  >> wrote:
> >
> > Hi,
> >
> > s anwar wrote:
> > > Please see attached source file.
> > >
> > > According to my understanding of MPI_Comm_spawn(), the
> > intercommunicator
> > > returned is the same as it is returned by
> MPI_Comm_get_parent() in
> > the
> > > spawned processes. I am assuming that there is one
> intercommunicator
> > > which contains all the (spawned) child processes as well as
> the parent
> > > process. If this is the case, then why does an MPI_Bcast()
> using
> > such an
> > > intercommunicator wait indefinately?
> >
> > your code from line 75:
> >
> > MPI_Bcast(_int, 1, MPI_INT, 0, intercomm);
> >
> > from the mpi 2 standard " 7.3.2.1 
> . Broadcast"
> >
> > "If comm is an intercommunicator, then the call involves all
> > processes in
> > the intercommunicator, but with one group (group A) defining
> the root
> > process. All processes in the other group (group B) pass the
same
> > value in
> > argument root, which is the rank of the root in group A. The
> root passes
> > the value MPI_ROOT in root. All other processes in group A
> pass the
> > value
> > MPI_PROC_NULL in root. Data is broadcast from the root to all
> > processes in
> > group B. The receive buffer arguments of the processes in
group B
> > must be
> > consistent with the send buffer argument of the root."
> >
> > so, you define no group to be the root group (group A). i
> don't know
> > what
> > schould happen, when no root group is defined, but your code
> firstly
> > don't
> > conform to the standard, so don't blame the implementation
first.
> >
> > by
> > bert wesarg
> >
> > ___
> > users mailing list
> > us...@open-mpi.org   us...@open-mpi.org >
> > http://www.open-mpi.org/mailman/listinfo.cgi/users
> >
> >
> >
> >
>

>
> >
> > ___
> > users mailing list
> > us...@open-mpi.org 
> > 

Re: [OMPI users] Why should the attached code wait on MPI_Bcast

[Bert Wesarg]

hi,

s anwar wrote:
> Thank you for the clarification. Why is MPI_COMM_SELF not the correct
> communicator for MPI_Comm_spawn(). My application will have a single
> master only.
yes for a single master this should be the same, but never try this.
> 
> Also, can I merge the intercommunicator into an intracommunicator via
> MPI_Intercomm_merge(intercomm, 0, ) and use MPI_Bcast(..., 0,
> intracomm) instead of sending MPI_ROOT or 0 (root's rank in local group).
dito, never try this, but it should work

bye
bert wesarg
> 
> Thanks.
> 
> On 7/18/06, *Bert Wesarg*  > wrote:
> 
> hi,
> 
> yes sorry for my first reply, my words were to rough.
> 
> a bcast for a intercomm works this way (in your words):
> 
> - your masters want to send a buffer to your slaves
> - one of the masters must provide the MPI_ROOT as root in the
> MPI_BCAST call
> - all slaves must provide the rank of this MPI_ROOT as the root argument
> - all others ranks from the masters, must provide MPI_PROC_NULL as root
> 
> so the buffer will be send from MPI_ROOT to all slaves
> 
> if this isn't that clear, you should understand the concept of mpi
> intercomms. a good starting point could be the standard:
> http://www.mpi-forum.org/docs/mpi-11-html/node111.html
> 
> 
> by the way:
> * your one_int is never intitialized
> * i don't know if the MPI_COMM_SELF is the right comm for the
> MPI_COMM_SPAWN
> 
> by
> bert wesarg
> 
> s anwar wrote:
> > I don't think I blamed the implementation in any way in my original
> > email. My intent is to gain understanding of why my code
> does/should not
> > work. I assumed that I was not passing the correct intercommunicator.
> > However, I am at a loss on how to construct a proper
> intercommunicator
> > in this case. You have mentioned that I haven't defined any group
> to be
> > the root group. Could you care to elaborate on how can I make a
> group a
> > root group?
> >
> > Thanks.
> >
> > On 7/18/06, *Bert Wesarg* < wes...@informatik.uni-halle.de
> 
> >  >> wrote:
> >
> > Hi,
> >
> > s anwar wrote:
> > > Please see attached source file.
> > >
> > > According to my understanding of MPI_Comm_spawn(), the
> > intercommunicator
> > > returned is the same as it is returned by
> MPI_Comm_get_parent() in
> > the
> > > spawned processes. I am assuming that there is one
> intercommunicator
> > > which contains all the (spawned) child processes as well as
> the parent
> > > process. If this is the case, then why does an MPI_Bcast()
> using
> > such an
> > > intercommunicator wait indefinately?
> >
> > your code from line 75:
> >
> > MPI_Bcast(_int, 1, MPI_INT, 0, intercomm);
> >
> > from the mpi 2 standard " 7.3.2.1 
> . Broadcast"
> >
> > "If comm is an intercommunicator, then the call involves all
> > processes in
> > the intercommunicator, but with one group (group A) defining
> the root
> > process. All processes in the other group (group B) pass the same
> > value in
> > argument root, which is the rank of the root in group A. The
> root passes
> > the value MPI_ROOT in root. All other processes in group A
> pass the
> > value
> > MPI_PROC_NULL in root. Data is broadcast from the root to all
> > processes in
> > group B. The receive buffer arguments of the processes in group B
> > must be
> > consistent with the send buffer argument of the root."
> >
> > so, you define no group to be the root group (group A). i
> don't know
> > what
> > schould happen, when no root group is defined, but your code
> firstly
> > don't
> > conform to the standard, so don't blame the implementation first.
> >
> > by
> > bert wesarg
> >
> > ___
> > users mailing list
> > us...@open-mpi.org   us...@open-mpi.org >
> > http://www.open-mpi.org/mailman/listinfo.cgi/users
> >
> >
> >
> >
> 
> 
> >
> > ___
> > users mailing list
> > us...@open-mpi.org 
> > http://www.open-mpi.org/mailman/listinfo.cgi/users
> 

Re: [OMPI users] Why should the attached code wait on MPI_Bcast

[Bert Wesarg]

hi,

yes sorry for my first reply, my words were to rough.

a bcast for a intercomm works this way (in your words):

- your masters want to send a buffer to your slaves
- one of the masters must provide the MPI_ROOT as root in the MPI_BCAST call
- all slaves must provide the rank of this MPI_ROOT as the root argument
- all others ranks from the masters, must provide MPI_PROC_NULL as root

so the buffer will be send from MPI_ROOT to all slaves

if this isn't that clear, you should understand the concept of mpi
intercomms. a good starting point could be the standard:
http://www.mpi-forum.org/docs/mpi-11-html/node111.html

by the way:
* your one_int is never intitialized
* i don't know if the MPI_COMM_SELF is the right comm for the MPI_COMM_SPAWN

by
bert wesarg

s anwar wrote:
> I don't think I blamed the implementation in any way in my original
> email. My intent is to gain understanding of why my code does/should not
> work. I assumed that I was not passing the correct intercommunicator.
> However, I am at a loss on how to construct a proper intercommunicator
> in this case. You have mentioned that I haven't defined any group to be
> the root group. Could you care to elaborate on how can I make a group a
> root group?
> 
> Thanks.
> 
> On 7/18/06, *Bert Wesarg*  > wrote:
> 
> Hi,
> 
> s anwar wrote:
> > Please see attached source file.
> >
> > According to my understanding of MPI_Comm_spawn(), the
> intercommunicator
> > returned is the same as it is returned by MPI_Comm_get_parent() in
> the
> > spawned processes. I am assuming that there is one intercommunicator
> > which contains all the (spawned) child processes as well as the parent
> > process. If this is the case, then why does an MPI_Bcast() using
> such an
> > intercommunicator wait indefinately?
> 
> your code from line 75:
> 
> MPI_Bcast(_int, 1, MPI_INT, 0, intercomm);
> 
> from the mpi 2 standard "7.3.2.1 . Broadcast"
> 
> "If comm is an intercommunicator, then the call involves all
> processes in
> the intercommunicator, but with one group (group A) defining the root
> process. All processes in the other group (group B) pass the same
> value in
> argument root, which is the rank of the root in group A. The root passes
> the value MPI_ROOT in root. All other processes in group A pass the
> value
> MPI_PROC_NULL in root. Data is broadcast from the root to all
> processes in
> group B. The receive buffer arguments of the processes in group B
> must be
> consistent with the send buffer argument of the root."
> 
> so, you define no group to be the root group (group A). i don't know
> what
> schould happen, when no root group is defined, but your code firstly
> don't
> conform to the standard, so don't blame the implementation first.
> 
> by
> bert wesarg
> 
> ___
> users mailing list
> us...@open-mpi.org 
> http://www.open-mpi.org/mailman/listinfo.cgi/users
> 
> 
> 
> 
> 
> ___
> users mailing list
> us...@open-mpi.org
> http://www.open-mpi.org/mailman/listinfo.cgi/users

Re: [OMPI users] Why should the attached code wait on MPI_Bcast

[Bert Wesarg]

Hi,

s anwar wrote:
> Please see attached source file.
> 
> According to my understanding of MPI_Comm_spawn(), the intercommunicator
> returned is the same as it is returned by MPI_Comm_get_parent() in the
> spawned processes. I am assuming that there is one intercommunicator
> which contains all the (spawned) child processes as well as the parent
> process. If this is the case, then why does an MPI_Bcast() using such an
> intercommunicator wait indefinately?

your code from line 75:

MPI_Bcast(_int, 1, MPI_INT, 0, intercomm);

from the mpi 2 standard "7.3.2.1. Broadcast"

"If comm is an intercommunicator, then the call involves all processes in
the intercommunicator, but with one group (group A) defining the root
process. All processes in the other group (group B) pass the same value in
argument root, which is the rank of the root in group A. The root passes
the value MPI_ROOT in root. All other processes in group A pass the value
MPI_PROC_NULL in root. Data is broadcast from the root to all processes in
group B. The receive buffer arguments of the processes in group B must be
consistent with the send buffer argument of the root."

so, you define no group to be the root group (group A). i don't know what
schould happen, when no root group is defined, but your code firstly don't
conform to the standard, so don't blame the implementation first.

by
bert wesarg