Can someone please explain the idea of "page-fault mechanism to propagate writes"?
I cannot fully catch this point.
hweetvpu
@anonymous I think this means explicitly marking the data (in the program) uncachable so that the accesses to them are "automatically" page faults. This will require the updates to be flushed to the memory directly, which means memory coherence won't be a problem.
kayvonf
Not uncachable, but rather as paged out. Like you said, this will trigger a page fault and the OS can intervene and issue the appropriate network communication to execute the access.
dmerigou
Coherence and consistency are also interesting problems in the case of distributed systems. For instance, a distributed file sytem like AFS implements « session semantics » meaning that changes to a file are visible only when the user that modified the file has closed it. In these settings, it is worth to tradeoff consistency for scalability and network usage.
Can someone please explain the idea of "page-fault mechanism to propagate writes"?
I cannot fully catch this point.
@anonymous I think this means explicitly marking the data (in the program) uncachable so that the accesses to them are "automatically" page faults. This will require the updates to be flushed to the memory directly, which means memory coherence won't be a problem.
Not uncachable, but rather as paged out. Like you said, this will trigger a page fault and the OS can intervene and issue the appropriate network communication to execute the access.
Coherence and consistency are also interesting problems in the case of distributed systems. For instance, a distributed file sytem like AFS implements « session semantics » meaning that changes to a file are visible only when the user that modified the file has closed it. In these settings, it is worth to tradeoff consistency for scalability and network usage.