Re: A proposal for a General Clustering Framework

["Bill Todd" <billtodd@foo.mv.com>]

----- Original Message -----
From: "Alan Robertson" <alanr@unix.sh>
To: "Bill Todd" <billtodd@foo.mv.com>
Cc: "David Brower" <David.Brower@oracle.com>; "linux-cluster"
<linux-cluster@nl.linux.org>
Sent: Tuesday, June 05, 2001 3:42 PM
Subject: Re: A proposal for a General Clustering Framework


> Bill Todd wrote:

...

> > and the lack of fine control
> > over marshalling activity (including the potential for [distributed]
> > resource-allocation deadlocks).
>
> I'd like for you to elaborate some more on these two items:

The comment was made in the context of a more general-purpose mechanism than
in fact seems to be under consideration here, and in particular in the
context of potentially medium-to-large messages (or portions of messages)
where at least the simpler credit-based communication mechanisms may not
work well (e.g., specific space may be pre-allocated at the receiver).

It's been many years since I thought about distributed resource allocation
deadlocks in detail, but in general they involve binary (or longer-cycle)
deadly embraces where backed-up request activity consumes the (usually
buffer) resources required to support shipment and/or receipt of the
responses that would free up the tied-up request resources.  One common
approach to the problem includes having requestors reserve response-receipt
space prior to sending requests and having responders always reserve
sufficient space to send at least one response (which is guaranteed not to
block, since the requestor has guaranteed the space to receive it).

The responder must be able to 'stream' larger responses through buffers of
limited size if the reserved response space requirement is to be bounded.
It would certainly be possible to create an RPC mechanism that would do
this, but it's not something I'd assume was there without checking first.

Sorry for the fuzziness above - as I said, it's been a long time.  And in
any event, with relatively small messages (in particular, with a known upper
size bound) the problem is much more tractable.

- bill



Linux-cluster: generic cluster infrastructure for Linux
Archive:       http://mail.nl.linux.org/linux-cluster/