I read a little more about MPI derived data types, and to answer my own question:
In general, we cannot assume portability of sending a C/C++ struct as a stream of bytes. There must be a promise that data representation on machines involved in the transmission must perform the same data layout.
Also, there was a +1 on using derived data types in terms of clarity of code.
But, now that I decided to use it, I run into another
I have a function that commits the new datatype: Add_New_MPITypes(). This is called just after MPI_Init(...).
After a few subsequent function calls, I am doing MPI-Send/Receive in another function, which looks like:
void sendMessagetoSlave(void* Payload, int MESSAGETYPE)
//Add_MPI_msgInstallP_Type(); /*Was already done in Add_New_MPITypes() */
InstallPMessage = *(msgInstallP*)Payload;
(void*)Payload, /* Payload */
sizeof(msgInstallP), /* size of the payload */
MPI_MSGINSTALLP, /* MPI Data type */
InstallPMessage.location, /* location to which the message is being sent */
MASTERSLAVECONTROLMESSAGE, /* Tag */
MPI_COMM_WORLD /* Communicator */
The linker complains that it does not know MPI_MSGINSTALLP derived datatype. Specifically,
the message from the linker is:
"‘MPI_MSGINSTALLP’ was not declared in this scope".
I have using mpic++ (1.4.2) to compile, and g++ (4.5.3) to link.
Can anyone help?
From: devendra rai <email@example.com>
To: Open MPI
Sent: Wednesday, 4 January 2012, 17:31
Subject: [OMPI users] Can we avoid derived datatypes?
I need to send a struct- datatype over MPI. Currently, I send the strcture as a series of MPI_BYTEs and on the other end, I dereference it as though it were a struct- type.
Something like this:
MPI_Ssend((void*)&MasterSlavePayload, sizeof(MasterSlavePayload), MPI_BYTE, destNode,MASTERSLAVECONTROLMESSAGE,MPI_COMM_WORLD);
where MasterSlavePayload is a structure variable.
This currently seems to work, where we have a homogenous environment: same hardware configuration, and same operating system.
The question is: Is this approach portable? Safe? And whether this will work on a
system of nodes with mixed processor types?
I read from MPI tutorials
"...Primitive data types are contiguous. Derived data types allow you to
specify non-contiguous data in a convenient manner and to treat it as
though it was contiguous. "
So, since I am using a primitive data type, does it mean that the packing of elements is maintained across the MPI_Send/MPI_Recv process? If so, it would mean that the approach that I use should work.
Thanks a lot,
users mailing list