Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!swrinde!zaphod.mps.ohio-state.edu!brutus.cs.uiuc.edu!apple!decwrl!sgi!vjs@rhyolite.wpd.sgi.com
From: vjs@rhyolite.wpd.sgi.com (Vernon Schryver)
Newsgroups: comp.protocols.nfs
Subject: Re: NeFS protocol
Summary: latency==latency
Message-ID: <61183@sgi.sgi.com>
Date: 30 May 90 06:44:04 GMT
References: <1990May24.034258.13625@Neon.Stanford.EDU> <3378@auspex.auspex.com> <23707@bellcore.bellcore.com>
Sender: vjs@rhyolite.wpd.sgi.com
Organization: Silicon Graphics, Inc., Mountain View, CA
Lines: 47

In article <23707@bellcore.bellcore.com>, mo@messy.bellcore.com (Michael O'Dell) writes:
> What might well be called "Clark's Law" is that "As bandwidth
> goes to infinity, latency becomes a constant 30 milliseconds."
> This is the round-trip delay across the US on a 1 gigabit or faster
> fibre-optic network.  You can't change that number unless you
> change the speed of light, so there is NOTHING you can do about
> latency except avoid round trips!!!
>
> The fact that some application can successfully beat the life out
> of a local ethernet cable is NOT a feature when considering 
> *networks which scale with latency*, and in Clark's view,
> and mine, them's the future of networking.   Everthing else
> is essentially a multiprocessor by comparison.
> 	-Mike

I am apologize for not being clear.

I claim doing 20,000 "RPC's" per second over ethernet (i.e. 50
microsecond/call) is in some sense equivalent to doing a more familiar
number (500/sec?) over a "high latency" medium such as cross-country
fiber.  In neither case is there time for one round trip per call.  One
round trip/call at 20,000 calls/sec comes out to 25 usec/packet.  Isn't
the minimum packet time, including preamble, something like 60.8 usec on
10MHz ethernet?

Instead of remote-file-system-programming-languages, some cheating seems
more to the point.   Obviously, to get 20,000 "RPC's"/sec, not to mention
the 1,000,000/sec that we would like, it was necessary to cheat.
Obviously, you can't let the application, whether it is the upper layers of
the file system or a graphics application, catch you cheating.

You can get impressive preformance gains many places by cheating.  The Sun
use of the directory bit for client cache control is effective chicanery.
The Silicon Graphics "private" NFS mount option is, IMH(&biased)O, even better.

Some of the newer network file systems show signs of the right kind of cheating.

Please note that 30 msec latency is not very bad.  It permits >=15 round
trips/second.  An NFS rsize/wsize of 64K would in principle give 1MB/sec of
data thru NFS, 5-10 times most current NFS implements.

The big deal with GByte (not Gbit) fiber is not the latency, but the
bandwidth*latency product.  Having MBytes in flight stresses lots of
interesting things.


Vernon Schryver        vjs@sgi.com