Re: Reliability of RPC services

[Jonathan S. Shapiro]

On Tue, 2006-04-25 at 23:54 +0200, Bas Wijnen wrote:

> Not exactly.  The problems of timeouts are limited to when they are used.
> That is, if we don't use timeouts in certain situations, then their problems
> don't hurt us in those situations.
> 
> This is different from single-copy capabilities, where the problems do impact
> situations when they aren't used.  I suppose this is the reason you don't like
> them.

Yes, but also for another reason. Single copy creates semantic
challenges, and it adds a measurable performance cost to support an
extremely rare error case. The same error case can be handled adequately
by watchdogs, which are required for any serious recovery handling in
any case (e.g. target simply stops, but does not get destroyed). Since
the watchdog must run anyway, the overhead in the kernel path is
providing exactly zero value.

shap