Some reasons:
* prefetches are not tracked at all. This only works for token
coherence
* Even so, if a demand miss occurs shortly after the prefetch is
issued, then L2 will issue a 2nd request instead of merging the demand
miss with the prefetch. Not a big deal in token coherence except for the
waste of bandwidth
* Under high contention, no priority is given to demand misses. Though
in the default configurations, there is little contention.
--Mike
> It is repeatedly noted that the prefetcher shown in your tutorial
> slides is not "the right way," and should not be used as a template, etc.
> What is wrong with it?
> _______________________________________________
> Gems-users mailing list
> Gems-users@xxxxxxxxxxx
> https://lists.cs.wisc.edu/mailman/listinfo/gems-users
>
|