Regarding the first bullet point, we are okay with broadcasting as a backup in the case that there are more than the maximum number of sharers because we can tell that we would only have to broadcast very rarely--from the cache invalidation pattern graphs, we can see that most invalidation only deals with a very low number of sharers, so broadcasting would only have to happen occasionally.
SR_94
Regarding the 3rd point - following the example from previous slide, if we have 10 pointers each 10 bit wide (for a total of 100 bits) then we can have maximum 10 sharers per line. So in case we exceed the max sharers and revert back to bit vector rep. scheme, each bit will correspond to ceiling(1024/100) nodes(Since we have a total of 1024 processors but only a 100 bits)?
shpeefps
Is there some kind of amortized analysis quantification that calculates how often the worst case scenario of broadcast is likely to happen and what is the overall improvement over just broadcasting?
yes
I think that would have to be measured empirically over multiple different types of programs to see what kind of improvement there would be, because depending on the number of sharers and frequency of access it could be very different.
Regarding the first bullet point, we are okay with broadcasting as a backup in the case that there are more than the maximum number of sharers because we can tell that we would only have to broadcast very rarely--from the cache invalidation pattern graphs, we can see that most invalidation only deals with a very low number of sharers, so broadcasting would only have to happen occasionally.
Regarding the 3rd point - following the example from previous slide, if we have 10 pointers each 10 bit wide (for a total of 100 bits) then we can have maximum 10 sharers per line. So in case we exceed the max sharers and revert back to bit vector rep. scheme, each bit will correspond to ceiling(1024/100) nodes(Since we have a total of 1024 processors but only a 100 bits)?
Is there some kind of amortized analysis quantification that calculates how often the worst case scenario of broadcast is likely to happen and what is the overall improvement over just broadcasting?
I think that would have to be measured empirically over multiple different types of programs to see what kind of improvement there would be, because depending on the number of sharers and frequency of access it could be very different.