Even on a single machine with multiple processors, I think that communication should happen here because the act of P1 writing into the row that P2 needs will dirty those cache entries, so that at some point in time, P2 will have to go to main memory or P1's cache to get the new data, instead of being able to use its own cache.
This comment was marked helpful 1 times.
black
I think P1 and P2 may not necessarily to be on the different cores, like hyper-threading. Then they just need to fetch the data from L3 shared-cache. Or, if they are just in different ALUs, then they can exchange data even faster.
This comment was marked helpful 0 times.
kayvonf
@black: you are completely right. However, in the context of this slide I was referring to the case where P0 corresponds to work done by processor 0, and P1 corresponds to processor 1. (I referred to processors when explaining the slide, but you can replace processor with core without any change to the argument.)
Even on a single machine with multiple processors, I think that communication should happen here because the act of P1 writing into the row that P2 needs will dirty those cache entries, so that at some point in time, P2 will have to go to main memory or P1's cache to get the new data, instead of being able to use its own cache.
This comment was marked helpful 1 times.
I think P1 and P2 may not necessarily to be on the different cores, like hyper-threading. Then they just need to fetch the data from L3 shared-cache. Or, if they are just in different ALUs, then they can exchange data even faster.
This comment was marked helpful 0 times.
@black: you are completely right. However, in the context of this slide I was referring to the case where P0 corresponds to work done by processor 0, and P1 corresponds to processor 1. (I referred to processors when explaining the slide, but you can replace processor with core without any change to the argument.)
@Q_Q. Good comment as well.
This comment was marked helpful 0 times.