Re: Minutes from Santa Fe DFS-WG meeting

"Drew D. Perkins" <> Thu, 05 December 1991 16:19 UTC

Received: from by NRI.NRI.Reston.VA.US id aa12981; 5 Dec 91 11:19 EST
Received: from by with SMTP; Thu, 05 Dec 91 11:15:01 -0500
Received: by (5.57/Ultrix3.0-C) id AA18436; Thu, 5 Dec 91 11:14:59 EST
Received: by (4.1/SMI-4.1) id AA19136; Thu, 5 Dec 91 11:02:17 EST
From: "Drew D. Perkins" <>
Cc: mdavies@NRI.Reston.VA.US
Date: Thu, 5 Dec 91 11:02:17 EST
Subject: Re: Minutes from Santa Fe DFS-WG meeting
Message-Id: <>

	From: Peter Honeyman <>
	Date: Thu, 5 Dec 91 00:06:06 EST
	Subject: Minutes from Santa Fe DFS-WG meeting

	NFS/TCP is commercially available, and is under development by many
	vendors.  Connection maintenance is not entirely a solved problem.

Four vendors demonstrated NFS over TCP at Interop Fall '91:
	FTP Software	- OS/2
	InterCon	- Macintosh
Other vendors with plans in this area include:
	FTP Software	- DOS
	Sun		- Solaris
	USL		- SVR4

I gave a talk on NFS over TCP at the Sun ONC/NFS conference '91.  I can make
copies of slides available if anyone is interested.  Topics included solutions
to connection maintenance as well as benefits of NFS/TCP vs. NFS/UDP,
performance, etc..

	Sun/RPC over TCP has problems with accurate RTT because the network
	latency is smeared by the upper-layer (i.e., NFS) service times.  (See
	"Transport Issues in the Network File System" by Bill Nowicki, Computer
	Communication Review 19(2), pp. 16-20 (April, 1989) for related work.)

I assume you meant Sun/RPC over UDP, not TCP.  Specifically, using UDP causes
the RTT to include disk queueing factors.  You are really measuring the RTT
to the disk, NOT the RTT of the network.  Tests so far indicate a very
significant (order of magnitude) reduction in retransmissions using our NFS/TCP
vs. NFS/UDP.

Drew Perkins		
INTERSTREAM, Inc.			email:
1501 Reedsdale St.			voice: (412) 323-8000
Pittsburgh, PA 15233-2329		fax:   (412) 323-1930