On Aug 6, 2012, at 9:36 AM, "Gong Ding" <gdiso at ustc.edu> wrote:
> Hi,
> I am trying to use AMS as data communicator between computational code and
> user monitor. The AMS works well for serial situation, however, I don't know
> how to use it for parallel environment. The manual has little information,
> too.
You need to build AMS with the same MPI that you will use in the parallel
program.
>
> I'd like to know some thing about data synchronization.
> Does AMS process data synchronization in MPI environment automatically?
> Should I call AMS_Comm_publish on every MPI processor, and create memory with
> AMS_Memory_create for my variables on each processor?
There are different types of variables you can publish.
AMS_Comm_publish() takes an argument AMS_Comm_type which would be MPI_TYPE
like in your case.
AMS_Memory_add_field() takes an argument AMS_Shared_type which can be
AMS_COMMON, AMS_DISTRIBUTED or AMS_REDUCED and if it is reduced the
AMS_Reduction_type is AMS_SUM, AMS_MAX, AMS_MIN, or AMS_NONE. See the manual
page for AMS_Memory_add_field() .
If AMS_Shared_type is AMS_COMMON then AMS assumes that the values in those
locations are identical on all processes (and it brings over to the accessor
one set of those values), if distributed then each process has an array of
different values (and it brings over all those values, in a long array, on the
accessor) if reduce that it brings back the values after applying the reduction
operator across all the processes; so with sum it beings back the sum of the
values over all the processes.
>
> If so, the accessor will see how many TCP/IP port? Or accessor only
> communicate with MPI processor 0 and processor 0 will broadcast the
> information to all the others?
AMS completely manages the fact there are multiple "publishers", the
accessor transparently handles getting stuff from all the publisher nodes.
>
> If not, should I only create AMS object on processor 0, and processor 0
> should broadcast what AMS get?
Nope, you do not need to do this.
Barry
>
> Gong Ding
>
>
>
>
>
>
>
