[gclist] Re: gclist-digest V1 #74

Matthew Fuchs matt@wdi.disney.com
Mon, 17 Jun 96 9:52:49 PDT


Marc:

> 
> >From: matt@wdi.disney.com (Matthew Fuchs)
> >Date: Fri, 14 Jun 96 10:07:45 PDT
> >Subject: Re: [gclist] Re: gclist-digest V1 #70
> >
> >The following was my response to Marc which shows why his scenario is
> >incorrect.  The basic point is that we assume the IRG is constantly
> >maintained, as per Marc's own work on SSP chains.
> 
> In SSP Chains, we only maintain an approximation of the remote inverse
> graph.  That is, each site knows which site might point to it, but not
> precisely which object holds the pointer.  Admittedly this information can
> be found, but not without a cost.  Furthermore, we do not maintain any
> local inverse graph information.  Doing so is hard.
> 
>                                         Marc

As this approximation always contains _at least_ the right edges, this
is actually sufficient.  If you assume we've built a local IRG, we
know which scions are part of which DGC operations, so whenever a new
incoming reference appears, we know to start distributed tracing along
that path.  This happens immediately and does not require any
additional search of the local graph.  The scion doesn't need to know
which exact objects refer to it - that's the job of the nodes
containing the stubs.  It just needs to know which nodes have stubs.
Of course, this throws all the weight (again) on building the local
IRG, which I'll discuss in your other email.

Matthew