You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Just posting this here, because I had an idea for a future feature of KaMPIng
I looked a bit into MPI_Bsend. The user may/has to attach a buffer allocated by himself via MPI_Buffer_attach. If a message may not be directly received upon send, it is buffered. This may be an alternative to Isend when sending many small messages. The problem with the current implementation is, that it may fail if the buffer is full. Maybe we can overcome this by blocking upon a buffer overflow (assuming that we get no deadlock) until it has been emptied, making the whole buffered sending process transparent to the user. It also seems like MPI uses a rather simple approach for buffering to be generically applicable, we should do some implementation studies and maybe could provide the user with alternative buffering strategies to plug in.
The text was updated successfully, but these errors were encountered:
Just posting this here, because I had an idea for a future feature of KaMPIng
I looked a bit into
MPI_Bsend
. The user may/has to attach a buffer allocated by himself viaMPI_Buffer_attach
. If a message may not be directly received upon send, it is buffered. This may be an alternative toIsend
when sending many small messages. The problem with the current implementation is, that it may fail if the buffer is full. Maybe we can overcome this by blocking upon a buffer overflow (assuming that we get no deadlock) until it has been emptied, making the whole buffered sending process transparent to the user. It also seems like MPI uses a rather simple approach for buffering to be generically applicable, we should do some implementation studies and maybe could provide the user with alternative buffering strategies to plug in.The text was updated successfully, but these errors were encountered: