Re: deferred deallocation of local objects

[Richard Frith-Macdonald]

PS. If we are to change the DO protocol, there is one optimisation I 
might like to add ..
piggyback proxy release messages on the end of message replies.

In the common case where a message has a an object argument passed by 
reference,
the current system has a network read to get the request, a write to 
send the response,
and then a write to release the argument.  If the release was part of 
the reply packet,
we would reduce this to two network write operations ... which might be 
a significant
performance gain.

The other (perhaps better) approach is to defer sending of release 
messages until
either a number have accumulated, or a timeout occurs, and send 
multiple releases
in a single packet (I already added support for multiple releases in a 
single packet).
This is not quite as elegant, but works for situations where all the 
messages being
sent are oneway-void so there are no reply packets.