I think the communication is performed en masse because there is a cost associated with sending messages, and sending by row instead of by element cuts down on the extra costs of message passing.
On that note, performance-wise, how does the message passing model compare to the other models? It would seem as though it would be slower.
This comment was marked helpful 0 times.
crs
@bstan I believe there are some VM tricks that can be applied to speed up message passing, but that is essentially shared address space model. I think message passing is slower compared to other models too as there may be lots of copying and synchronization. Wonder if there is any special hardware support for this kind of model.
This comment was marked helpful 0 times.
Q_Q
I think an example of message passing using VM tricks is to mmap a region of memory, write into it, then somehow tell another process to also map that memory, so it can read/write to it. That does sound a lot like shared address space though...
I suppose this could also be done with pipes, but in that case the OS is probably doing copying between buffers.
This comment was marked helpful 0 times.
ron
As another perspective, Message Passing is designed as a way to facilitate parallel processing on heterogeneous distributed systems with distributed memory. Applications that use it are probably designed to have minimal use of shared data and do most of their computation on local memory, with communication via small-sized sends/receives over the network. It doesn't seem like you can improve upon this model by implementing a shared address space, since that really is just a layer of abstraction, and would require additional synchronization that would still at the low level be implemented by point-to-point sends and receives.
This comment was marked helpful 0 times.
shabnam
I am sure message passing has some overhead in establishing a connection before actual data is exchanged. In this we want to minimize that overhead by doing communication en masse.
I think the communication is performed en masse because there is a cost associated with sending messages, and sending by row instead of by element cuts down on the extra costs of message passing.
On that note, performance-wise, how does the message passing model compare to the other models? It would seem as though it would be slower.
This comment was marked helpful 0 times.
@bstan I believe there are some VM tricks that can be applied to speed up message passing, but that is essentially shared address space model. I think message passing is slower compared to other models too as there may be lots of copying and synchronization. Wonder if there is any special hardware support for this kind of model.
This comment was marked helpful 0 times.
I think an example of message passing using VM tricks is to mmap a region of memory, write into it, then somehow tell another process to also map that memory, so it can read/write to it. That does sound a lot like shared address space though...
I suppose this could also be done with pipes, but in that case the OS is probably doing copying between buffers.
This comment was marked helpful 0 times.
As another perspective, Message Passing is designed as a way to facilitate parallel processing on heterogeneous distributed systems with distributed memory. Applications that use it are probably designed to have minimal use of shared data and do most of their computation on local memory, with communication via small-sized sends/receives over the network. It doesn't seem like you can improve upon this model by implementing a shared address space, since that really is just a layer of abstraction, and would require additional synchronization that would still at the low level be implemented by point-to-point sends and receives.
This comment was marked helpful 0 times.
I am sure message passing has some overhead in establishing a connection before actual data is exchanged. In this we want to minimize that overhead by doing communication en masse.
This comment was marked helpful 0 times.