SUMMARY: NFS over 64Kbps data circuit

From: Paul Begley (
Date: Fri Mar 05 1993 - 13:31:58 CST

Sorry for the late posting! This is the summary of replies I
received regarding the viability of using NFS over an international 64K
data circuit.

As always, the sun-managers have been very responsive! Due to the
large response, the list of contributors is appended to this posting.

I summarized the comments and included any published performance or
parameter information people included. Related papers and some non-NFS
software is included at the end of the summary. Due to the problems of
coordinating this project between four companies we have not
implemented NFS yet. If there are startling revelations, I will post
an additional summary, particularly after we have a network manager
installed and can tabulate performance data on the circuit.

1. NFS over a WAN is possible. Many people are running NFS over
56K/64K lines on production systems. Several people are using 19.2K
lines (including SLIP), but performance is not the best. Several
people did not find NFS performance satisfactory for production work,
but from the mail I received, it is function of the type of NFS traffic
you expect. People using automounter (multiple mounts) and performing
NFS writes suffered the most. Large file transfers with NFS do not
seem to work (>10MB).

Another alternative suggested in lieu of NFS was using ftp and rdist
with crontab to coordinate files between sites. This was of particular
interest with international sites where the time differences cause our
work hours to only overlap by 3-4 hours or so.

Rules of thumb for actual file transfers between desktop SPARCs (from
Ron Vasey):
    ethernet LAN: 8mbytes/min (rsh pipeline, rdump, ftp)
                     5mbytes/min (when READING NFS mount, eg cp)
    ISDN/ethernet: 300kbytes/min (ftp) -- 2x64kb channels
    (so 64K channel would be ~150kbytes/min)

Example mount parameters for 2MB fiber X.25 link (from Michael Stumpf):
mount -o rw,bg,retry=100,rsize=512,wsize=512,timeo=90,retrans=9,soft,intr \
  cshomehost:/home /home/cogsys

2. NFS configuration should be modified to decrease buffer size
and increase timeout interval. Turning on UDP checksum was also

TCP/IP header compression is available with Cisco routers and probably
others. We will be experimenting with this and as soon as we get a
network manager installed, we should be able to evaluate any
improvement in performance this affords.

3. Read Hal Stern's book 'Managing NFS and NIS'

4. NFS reads are OK, some people experienced data loss with NFS
writes. The data transfer differences between NFS reads and writes
are well documented, so this is not a shock.

5. ping time from Ireland to Philadelphia, PA is about 125 ms
ftp speed is about 8 kbytes/sec with a moderate system load between a
Sun IPC and an IBM RS/6000.

Other users reported ping times of up to 200 ms and ftp speeds of

Related Papers: Thanks to Hal and Christian for these included papers.
>From Hal Stern - Hal forwarded an article he wrote for the /sys.admin
column, SunWorld, July 1992

>From Christian Lawrence - attached a paper on UDP/RPC tuning which
was very helpful

Related Software: WWFS Cluster Server (CS)
Youki Kadobayashi
WWFS Research Group
Department of Information and Computer Sciences
Osaka University, Toyonaka 560, Osaka, Japan

WWFS Research Group at Osaka University is happy to announce the first
release of the WWFS Cluster Server (CS) -- the Internet file cache, an
intermediary file server, an NFS-FTP gateway software, or an user-level
NFS server, that runs on top of vanilla BSD.

To run this release, you need either a Sun, Sony NEWS, Omron LUNA, or MIPS
as a platform. You need "perl". No kernel modification is necessary.

We hope you enjoy this system; more information about the CS is
available in the documentation directory of the CS distribution.
Please direct any questions to
WWFS mailing list <>
or, me <>.

- You can mount existing anonymous ftp servers.
- You can forget details about hostnames and directories.
- You can forget which files you've transferred already.
- You can share files with other people in your cluster.

Main shortcomings
- No automatic server selection yet.
- Volumes are described in locally maintained files.
- Partial support for CS-specific or FTP-specific semantics.
- No multi-level caching yet.

How to get related information?
For more information, just grab the distribution "wwcs-1.0.tar.Z" from
your nearest archive, possibly: (anonymous ftp) (anonymous ftp)
        <insert your own archive here>

Technical Reports from the WWFS Research Group at Osaka University is
also available via anonymous ftp from:

A big thanks to all who responded:
Mr T Crummey (DIJ) <> (David Galloway)
Patrick O'Callaghan <>
Christian Lawrence <> (Alan Medsker)
Glenn Shaw <>
aldrich@sunrise.Stanford.EDU (Jeff Aldrich)
Art.Hebert@EBay.Sun.COM (Art Hebert) (Dick St.Peters)
Dieter Muller <> (Rex Espiritu) (Mike Andrews)
stern@sunne.East.Sun.COM (Hal Stern - NE Area Systems Engineer) (Brett Lymn)
Phil Race <>
David Fetrow <>
eclipse! (Jae Chung) (Jeff Leader)
ups!kalli!kevin@fourx.Aus.Sun.COM (Kevin Sheehan {Consulting Poster Child})
tricia@Erin.GOV.AU (Tricia Kaye)
ups!upstage!glenn@fourx.Aus.Sun.COM (Glenn Satchell-Uniq Professional Services)
Rodney Campbell <>
Mike Raffety <> (Michael Stumpf)
ger barrett <>
Youki Kadobayashi <>

Paul Begley                            Internet:
United Engineers & Constructors        Voice:     011-353-21-378811
Ringaskiddy, Ireland                   FAX:       011-353-21-378285  

This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:07:33 CDT