Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!seismo!rutgers!mit-eddie!bacchus!wesommer
From: wesommer@athena.mit.edu (William Sommerfeld)
Newsgroups: comp.unix.questions
Subject: Re: Dreams of NSF(A) and ....
Message-ID: <456@bacchus.MIT.EDU>
Date: Sun, 19-Apr-87 17:32:15 EST
Article-I.D.: bacchus.456
Posted: Sun Apr 19 17:32:15 1987
Date-Received: Sun, 19-Apr-87 22:43:01 EST
References: <3680010@nucsrl.UUCP>
Sender: daemon@bacchus.MIT.EDU
Reply-To: wesommer@athena.mit.edu (William Sommerfeld)
Organization: MIT Project Athena
Lines: 31

In article <3680010@nucsrl.UUCP> ram@nucsrl.UUCP (Renu Raman) writes:
>     Has anybody tried NFS over long haul networks? (Something like ARPA,
>     dedicated High speed links over large geographical distances....).

Yes.  It doesn't work very well over a long haul ARPA link (say,
MIT->UCB).  It probably works OK over T1 links (like the NSFnet
backbone).

NFS was designed for high-speed, low-latency, high-reliability nets
like Ethernet.  On long-haul nets, two things kill NFS: the round trip
time, and the maximum packet size.  Every pathname *component* lookup
takes one exchange with the server.  On a local ethernet, this takes
on the order of a few milliseconds (~10).  Over the ARPAnet, which is
made of 56KB links for the long haul, 1 *second* is more like reality.
This means that looking up, say /usr/spool/news/net/unix/wizards/400
on UCBVAX would take 7 roundtrips, probably taking seven or eight
seconds.  Then, after that, each read, of no more than perhaps 450
bytes, would take another second.  (NFS returns about 70 bytes of file
statistics information with each read request, given that the maximum
packet size on the ARPAnet is 576 bytes, and the UDP/IP headers
themselves take perhaps 40 bytes).

That's a throughput of 3.2KBits/second on a 56KB/sec link - about 6%
of the bandwidth theoretically possible.  According to SUN, NFS over
ethernet between a pair of SUN-3's moves more like 250 KBytes/sec
(2MBit/sec), which is 20% of Ethernet speed.

					Bill Sommerfeld
					MIT Project Athena
				  ARPA: wesommer@athena.mit.edu
				  UUCP: ...!mit-eddie!wesommer