Path: utzoo!utgpu!news-server.csri.toronto.edu!mailrus!cornell!ken
From: ken@gvax.cs.cornell.edu (Ken Birman)
Newsgroups: comp.sys.isis
Subject: Re: How to find rank of server w.r.t. bcast messages
Message-ID: <44143@cornell.UUCP>
Date: 4 Aug 90 00:53:10 GMT
References: <219@newhope.water.ca.gov>
Sender: nobody@cornell.UUCP
Reply-To: ken@gvax.cs.cornell.edu (Ken Birman)
Organization: Cornell Univ. CS Dept, Ithaca NY
Lines: 39

In article <219@newhope.water.ca.gov> rfinch@caldwr.water.ca.gov
	(Ralph Finch) writes:
>
>We wish to more or less use the parallel computing concept as outlined
>in the Isis manual (the fft example), but instead of rank in the group
>we wish to use rank according to the order the servers receive the
>bcast message.  Is this possible in an easy fashion (Isis call)?


If I understand the question, yes.  You will need to use abcast to send
to the servers, or use cbcast and have some "primary" server that assigns
rankings and periodically tells the others with a cbcast so that they know
what it picked.  As a rule of thumb, abcast is about half as fast as 
cbcast, so this second scheme is only worthwhile if you get such a high
volume of requests that a little delay before sending the ranking won't
impact performance.

Assuming that you just use abcast, the servers will see the requests in
the same order and will get them in the same "state" (i.e. group view, etc)
They can count requests and use this number as a ranking on the requests.
Everything else would be as in the example from the manual (well, maybe
it would be a good idea to count modulo the number of servers, if you
want the "ranking" to be in the range 0..nservers-1!)

If you use the cbcast scheme to reduce costs, the problem is that the
recipients won't see requests in the same order and hence can't use any
local rule to rank them.  This is why you would need a primary "ranker"
who would have to tell the others what ranking it picked for each request.

By the way, this is precisely how abcast works.  The abcast "ranking"
is just the order in which the messages should be delivered.

Life gets more complex if you want to do a uniform load-sharing scheme.
I assume this isn't quite what you have in mind, so I'll leave this as 
a puzzle for readers... (there seems to be a research topic here: although
there are many ways to solve this problem, no single solution jumps out as
obviously the one to use).

Ken