As Kayvon mentioned in lecture, in practice, for most datasets, only a small number of processors are sharing a particular cache line. Because of this, there is a significant advantage to using point-to-point communication versus broadcast communication for cache updates. For example, we may only need to communicate with 2 other sharers instead of all 63 other processors for broadcast.
This comment was marked helpful 0 times.
aew
As the next slide mentions, there is a small percentage of lines that are shared among many processors, such as the root node in Barnes Hut. This can be seen at the very right-most bucket of each histogram. However, these lines do not tend to present an issue to directory-based cache coherence performance, because they are not frequently written to, and instead mostly read.
As Kayvon mentioned in lecture, in practice, for most datasets, only a small number of processors are sharing a particular cache line. Because of this, there is a significant advantage to using point-to-point communication versus broadcast communication for cache updates. For example, we may only need to communicate with 2 other sharers instead of all 63 other processors for broadcast.
This comment was marked helpful 0 times.
As the next slide mentions, there is a small percentage of lines that are shared among many processors, such as the root node in Barnes Hut. This can be seen at the very right-most bucket of each histogram. However, these lines do not tend to present an issue to directory-based cache coherence performance, because they are not frequently written to, and instead mostly read.
This comment was marked helpful 0 times.