backwards compatibity, etc

New Message Reply About this list Date view Thread view Subject view Author view Attachment view

From: rick@snowhite.cis.uoguelph.ca
Date: 05/22/98-12:45:10 PM Z


From: rick@snowhite.cis.uoguelph.ca
Date: Fri, 22 May 1998 13:45:10 -0400 (EDT)
Message-Id: <199805221745.NAA11993@snowhite.cis.uoguelph.ca>
Subject: backwards compatibity, etc

Well, here's my $0.01 worth...

I think that V4 should try and address some new ground, but should not try
and solve all the world's distributed file system problems. (My sentiments
are along the lines of what Dave said, which was a good post imho.)

A few random thoughts:
- As Dave mentioned, the major stumbling block for replication is the
	file handle, so I think serious thought should be put to alternative
	ways to identify a file (this also addresses the issue of location
	independence, or at least looks at it). At this point, I would
	lean towards a longhand identifier (such as a URL), with a short
	hand form similar to a file handle, but I haven't really thought
	this through.

- If the file system is going to perform well over long haul nets, it must
	get away from the fine grained interaction typical of NFS RPCs.
	I really think that this comes down to some server state, such as
	"host A has attributes for file X cached", so that the client can
	afford to do long term caching without repeatedly going "are file
	X's attributes still YYY?". What I'm talking about here is along
	the lines of a server -> client callback system for updating
	client information, based upon server state. (I know, "NFS is
	stateless and that keeps recovery simple", but at some point
	this has to be questioned again.)

- I'll go one step farther and suggest that V4 should be a protocol
	designed to work on top of TCP only. My argument for this is that
	you can easily tie server state to a TCP connection and server
	state recover is done whenever a TCP connection is re-established
	after it is broken, so the state recovery dilemma falls out
	pretty easily. I don't see an argument for using UDP any longer
	since NFS over TCP is now well enough understood and performs
	as well as UDP in most cases.

It might be interesting for the group to look at the log of an old email
list that Dr. Jeff Mogul had going a long time ago on state recovery and
similar issues for NFS.

Anyhow, keep the disscussion going, rick


New Message Reply About this list Date view Thread view Subject view Author view Attachment view

This archive was generated by hypermail 2.1.2 : 03/04/05-01:45:45 AM Z CST