Path: utzoo!utgpu!news-server.csri.toronto.edu!cs.utexas.edu!swrinde!zaphod.mps.ohio-state.edu!think.com!barmar
From: barmar@think.com (Barry Margolin)
Newsgroups: comp.protocols.tcp-ip
Subject: Re: Is the Internet usable for wide-area interactive conversations?
Message-ID: <1991Jun18.224350.21721@Think.COM>
Date: 18 Jun 91 22:43:50 GMT
References: <2039.Jun1803.33.1391@kramden.acf.nyu.edu>
Sender: news@Think.COM
Reply-To: barmar@think.com
Organization: Thinking Machines Corporation, Cambridge MA, USA
Lines: 41

In article <2039.Jun1803.33.1391@kramden.acf.nyu.edu> brnstnd@kramden.acf.nyu.edu (Dan Bernstein) writes:
>And how good was that IP service? The minimum round trip time was rather
>impressive: one sixth of a second. But the maximum (not counting the
>unreachable periods) was awful: over ten seconds. The average was over a
>second, and the sample standard deviation was a whopping 1.8 seconds.

Sounds like a problem at the NYU end.  I tried pinging Berkeley.EDU from
here for a few minutes, and my maximum time is closer to your your minimum.
My times were: minimum 90ms, maximum 370ms, average 120ms.  And we're 18
hops away (7 hops across NEARnet, 7 hops across the NSFnet T3 backbone, and
4 hops across BARRnet and Berkeley's internal network).

I sent one packet every five seconds for ten minutes, and had 5% packet
loss.  I'm now trying a similar test using UDP, to see if I can elicit any
ICMP errors (ping doesn't notice down routes, because we use a default
route for everything outside the TMC local network, and routers don't
generate ICMP errors for ICMP packets).  I ran it for about five minutes
(again, 1/5 hz) without a single error (the UDP implementation I'm using
performs automatic retransmission for query/response protocols, so it would
only signal an error for several lost packets in a row or ICMP errors).

>That means you'd expect one packet in every few hundred to have a delay
>of seven seconds or more. Many studies have shown that people work best
>with a consistent feedback delay---it's much more important that the
>standard deviation be small than that the minimum be small.

I haven't bothered computing the std.dev. of my times.  A casual glance at
them (since I only pinged every five seconds, my sample size is easy to
look at manually) shows that most are in the 100-120ms range.  I think
people can live with that.

>I see this behavior all the
>time, even on links as short as BNL to NYU.

More evidence that the problem is somewhere near NYU.  Other posters have
mentioned that Nysernet has been having problems lately.
-- 
Barry Margolin, Thinking Machines Corp.

barmar@think.com
{uunet,harvard}!think!barmar