Rethinking Persistence in DIMSUM
16 Jan 95 15:04:00 +1600
>it seems necessary to introduce ... Fail-fast support: given the
sudden unavailability of a page due to
(e.g.) a machine crash, a mechanism is needed to trace back to all
processes that map that page, including those that are not
currently running so as to notify them.
Why is it that having a page MAPPED should qualify me for failure
notification? Carried to its logical conclusion, fail-fast would say
that if I have a key to (a segment containing) the page, I should be
notified (even if I don't have it mapped currently). Or if I could get
such a key from a directory I have, ...
The other logical extreme (fail-slow), which is what KeyKOS did (at
least in theory), is to notify only when a process attempts to actually
reference the page.
Do you have any arguments for the necessity of fail-fast, or why having
a page mapped is the right criterion for notification?
I have to disclose that Tandem does the equivalent of what you propose,
and I'm not sure it really makes sense there either.