[cap-talk] Communication successes (was: Re: Session failures)

[Jonathan S. Shapiro]

On Sat, 2008-03-15 at 11:26 -0700, Jed Donnelley wrote:
> I think if I
> come to understand the workings of the Three Party Commit
> protocol, which of course has to deal with the issues of
> lost messages that the Two Generals face, then I'll be in a
> better position to understand other efforts at distributed
> ... "consistency"?
> 
> What would you say your checkpoint/restart mechanism achieves
> Jonathan?

Assume, for a moment that a perfectly reliable N-way commit could be
achieved (it can't, but it doesn't matter).

The entire motivation for my attempt emerged from recognizing that if
N-way commit were possible, no real-world implementation of such a
protocol could be acceptably realized on any current or foreseeable
network technology. The metric of "acceptable" is that the total "halt
for agreement" interval must be no larger than 2ms. Because of light
speed limits and software protocol processing overheads, it is not
possible to make even a single round trip message within a 2ms bound on
a *long-haul* network.

It is well established that all of the N-way commit protocols require
multiple rounds. Therefore, an acceptable solution to distributed
checkpoint in the face of the 2ms limit must not rely on N-way commit,
or even on two-party synchronous agreement.

What I think my mechanism achieves is convergence on consistency within
bounded time and with bounded rollback *without* synchronous agreement
about when to take snapshots.

I can and will dig out the tech report and make it available. As I
mentioned earlier, I probably won't get to that before the 25th.

As background: the 2ms limit was motivated by human perception
thresholds for audio artifacts. For reference, the human adrenal reflex
path is a bit slower than this, but 2ms falls within the capabilities of
the noradrenal reflex path. It is reliably possible for humans to
trigger successive volitional muscle activity in sub-millisecond
intervals.

Aside: sequential behaviors can be trained into the noradrenal paths and
can subsequently be triggered volitionally. It's actually quite
astonishing how much the noradrenal path can be trained to do, and how
delicate its trained actions can be made while still operating at full
speed.

> ... I have ... reason to carefully consider Jonathan's thoughts 
>and writing ;-)

God have pity on your poor, demented soul.


shap