[moca] RE: Re: RE: RE : synchronous implementation of Pi calculus with choice?

[Fabrice Le Fessant <fabrice@xxxxxxxxxxxxx>]

>  I hope this does not sound to naive , but -
>  If all your process have transactioning and can be restored in
>  the event of failure and if you have a guranteed mesage delivery
>  system , both of which I belive are possible , why does the
>  possiblity of message failure matter for an algorithm that solves
>  the global consensus problem , if there was such an algorithm ?

If you have a way to restore the state after a failure, then there are
no failures: a crash can be seen simply as a slow down, and the host
will simply delay receiving and sending of messages until the failure
has been recovered. But this means that your restoration system cannot
fail, which is still a bit unlikely.

And guaranteed message delivery is impossible to implement too, if you
want to be able to limit the size of the buffers used for
messages. Indeed, you have either to remember which message you have
sent (if the message has not been received and need to be resent), or
the acknowledgement you have sent (if it has not been received and
need resent), or the acknowledgement of the acknowledgement, and so
on, so that you finally need at least one buffer slot per message.  Of
course, you can try to optimize by grouping messages or acks (or using
views of group-membership), but the problem still remains if the
number of hosts is not bounded.

- Fabrice
  
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
The "models for mobility" mailing list     mailto:moca@xxxxxxxxxxxxxxx
 http://www-sop.inria.fr/mimosa/personnel/Davide.Sangiorgi/moca.html