Initial filesystem design synopsis.
dillon at apollo.backplane.com
Thu Feb 22 10:37:06 PST 2007
:At 22:22 21/02/2007, Matthew Dillon wrote:
:> Plus, I need a name for this baby. I can't use DFS, however much I
:> want to, because the term is already over-used.
:If you went for 128-bit object IDs you could call it UFSv6 :-) :-)
:But seriously, instances of this filesystem will potentially be very
:big, very active *and* have very long lifetimes - is 64 bits really enough?
:Bob Bishop +44 (0)118 940 1243
:rb at gid.co.uk fax +44 (0)118 940 1295
What would be a reasonable limit of the number of replication targets
operating in master mode (master == changes can be made, slave == changes
can only be brought in through replication protocols).
Lets choose 64. Now, what would be a reasonable limit for the number of
discrete MODIFYING operations per second? Lets choose 10000000
So, that's 6 bits to identify replication master, and 24 bits for
sub-second resolution, leaving 40 bits for one-second resolution.
40 bits of one-second resolution comes to around 34865 years.
And its even better then that. For this filesystem, transaction ids
are created by synchronization events for modified data, not modifying
events. Such events occur far less often then the actual modifying
operations execuuted by programs.
64 bits is enough.
<dillon at backplane.com>
More information about the Kernel