[AFS3-std] Re: A call for consensus on draft-deason-afs3-type-time-02

[Tom Keiser]

On Mon, Aug 1, 2011 at 1:53 PM, Andrew Deason <adeason@sinenomine.net> wrote:
> On Mon, 1 Aug 2011 13:32:46 -0400
> Derrick Brashear <shadow@gmail.com> wrote:
>
>> On Mon, Aug 1, 2011 at 1:28 PM, Andrew Deason <adeason@sinenomine.net> wrote:
>>
>> >> b) the granularity
>> >
>> > This one I still have no idea on. I see reasons for both sides.
>>
>> So is there a reason an extended union with the various stamp
>> granularities would be a nonstarter? In particular I'd suggest the
>> draft strongly discourage
>> sending a larger timestamp than actual information supports (e.g.
>> don't use bits to send precision you don't have, rather than
>> trailing-zero-padding a
>> larger than needed number)
>
> Well, the objection to just having 64-bit seconds and 32-bit nanoseconds
> is "space", and a union tag is an extra 32 bits... If we had a "100 NS
> granularity" tag, then we'd have 100-ns granularity in 96 bits, whereas
> now we could have 1-ns granularity in 96 bits. Unless there's some other
> scheme you're thinking of that somehow makes this more efficient?
>

This is why in Pittsburgh we proposed that overarching structures
(e.g., AFSFetchStatus, and AFSStoreStatus) should be wrapped in
ext-union, while proscribing the use of union/ext-union for individual
members of those structures.  As I suggested last Friday, we could
define two different AFSFetchStatus/AFSStoreStatus implied legs for
now: one with 64-bit time, and a second with 96-bit time (perhaps, we
would also want to define a third leg with 32-bit time--given that
we're likely to have implementations of the protocol long before
server backing stores can handle the increased resolution)...

-Tom