Previous | Next --- Slide 18 of 36
Back to Lecture Thumbnails
sss

As all memory coherence operations are broadcasted to all processors, wouldn't this put a lot of pressure on the interconnect? Especially in the case that the interconnect is not a connected graph structure, but instead is the token based structure. Which is the case when there's in the range of 16 or more processor in a core.

paraU

I'm also curious about the interconnect. Will there be conflicts when multiple cores are shouting?

kayvonf

Absolutely. Contention for the interconnect can become quite a problem in snooping-based protocols. Hence the need for directory-based schemes that replace broadcast with point-to-point messages among only the interested caches.

uhkiv

How do directory-based schemes and snooping-based schemes compare in real life? Which one is more commonly used solution or does it depend on the case (i.e. above n number of cores the advantages of directory-based schemes overpower the advantages of snooping-based schemes)?

vrkrishn

I remember hearing that a lot of current systems have space for directories built into the hardware for the machine, so I imagine the directory based system has its merits. Not having to send signals to all processors about cache coherence must be significant enough of a benefit that modern system designers choose to implement a directory-based cache coherence.

Also look at slide 21 on the directory based cache scheme slide. In that slide we see that most programs only share data between a couple of processors at any given time. Although broadcasting might be efficient if we only have a couple of processors, as systems incorporate more cores the lack of complexity of broadcasting starts to get outweighed by the fact that signals are being sent to processors that don't need signals