I'm a little confused about the purpose of ghost cells. Are they just cells that aid communication between the different partitions? Would they be the cells along the boundaries?
This comment was marked helpful 0 times.
kayvonf
Remember this implementation is for a distributed address space system (e.g., a cluster of machines). "Ghost cells" are cells that are replicas of cells "owned" by another processor. The ghost cells store a replicated copy of the data on the local processor because this data is a necessary input to computations that actually update data owned by the local processor. Here, consider how in the processor of updating face 6, the update computation might need information from face 15.
In the diagram it's best to consider that all the grey cells, both dark and light are owned by processor P, and that the white ghost cells are owned by remote processors.
This comment was marked helpful 0 times.
benchoi
Am I right then in saying that ghost cells are like non-exclusive copies of a piece of memory held in the caches of processors other than the one actually performing operations on it?
This comment was marked helpful 0 times.
vrkrishn
I think the key point is that the data is the ghost cells is only brought over to another processor because they are necessary to the computation of results on a processor. The only reason that a processor would suffer the communication cost of bring cells over is if the cells are needed for computation
This comment was marked helpful 0 times.
kayvonf
@benchoi: sure, that's a good way to think about it. However, you should understand this is being done by the application, not the hardware. It is not replication of a memory address (e.g., like how a cache transparently replicated data in the address space), it's the application making copies of the data on its own.
I'm a little confused about the purpose of ghost cells. Are they just cells that aid communication between the different partitions? Would they be the cells along the boundaries?
This comment was marked helpful 0 times.
Remember this implementation is for a distributed address space system (e.g., a cluster of machines). "Ghost cells" are cells that are replicas of cells "owned" by another processor. The ghost cells store a replicated copy of the data on the local processor because this data is a necessary input to computations that actually update data owned by the local processor. Here, consider how in the processor of updating face 6, the update computation might need information from face 15.
In the diagram it's best to consider that all the grey cells, both dark and light are owned by processor P, and that the white ghost cells are owned by remote processors.
This comment was marked helpful 0 times.
Am I right then in saying that ghost cells are like non-exclusive copies of a piece of memory held in the caches of processors other than the one actually performing operations on it?
This comment was marked helpful 0 times.
I think the key point is that the data is the ghost cells is only brought over to another processor because they are necessary to the computation of results on a processor. The only reason that a processor would suffer the communication cost of bring cells over is if the cells are needed for computation
This comment was marked helpful 0 times.
@benchoi: sure, that's a good way to think about it. However, you should understand this is being done by the application, not the hardware. It is not replication of a memory address (e.g., like how a cache transparently replicated data in the address space), it's the application making copies of the data on its own.
This comment was marked helpful 0 times.