[35] in Commercialization & Privatization of the Internet
Operational use at capacity is useful sooner rather than later.
daemon@ATHENA.MIT.EDU (John Gilmore)
Sun Oct 28 22:05:27 1990
Date: Sun, 28 Oct 90 15:59:33 PST
From: gnu@toad.com (John Gilmore)
To: com-priv@psi.com
>From: Stephen Wolff <steve@cise.nsf.gov>
> Not the NSFNET Backbone. We see average loadings over 10% routinely, and
> peak loads over 70%.
WHAT?! The NSFnet is running at 70% capacity AT PEAK!? It's not turning
back ANY traffic for lack of bandwidth? It's not even asking TCP to throttle
back (by dropping packets), which modern TCP's do gracefully when asked??
Yeah, let's spend LOTS more money on capacity!
If NSF really does want to further the development of widespread
computer networks, as is alleged by various people in this group, the
worst thing they could do is put off the next round of protocol
performance tuning! We *really* don't want to hit the wall and crash
at 45Mbps, after increasing the number of users, the amount of data in
motion, and the impact on education, reasearch, industry, and private
parties by a factor of 30! The severe performance problems noticed
before the T1 backbone would not have been solved if we'd just thrown
capacity at the problem. There are still a lot of locally-optimal
globally-pessimal algorithms in our network today, that won't let it
scale to the sizes we're shooting for. Remember when the phones didn't
work during last year's SF earthquake? Let's find out now, rather than
at some critical time when the network fails millions, not hundreds
of thousands, of people.
> Moreover, NFS is not as large a fraction of the Backbone traffic as it is
> typically on local- and campus-area nets; this says to me that the new
> paradigm of computing -=> which you guys invented <=- hasn't yet spread
> beyond the campus - but when it does (and it will: e.g., I hear the Fermilab
> folk want to remote-mount the L3 data at CERN), look out!
NFS gets little use nationwide or worldwide because its performance
sucks. Berkeley CSRG has reimplemented NFS and put it on TCP, where it
belongs, so it can use all the congestion-avoidance built in to TCP.
Sun tried to back-patch these algorithms into their UDP NFS, but the
lack of wide-area use bespeaks failure. The Berkeley implementation
will probably be available next year, so you can expect more NFS
traffic then -- but it will be well-behaved NFS traffic, and will share
the bandwidth as equally as it knows how, with all the other traffic on
the network.
Note that T3 lines can't make UDP NFS any faster -- they don't reduce
latency, just increase capacity. Your compiles doesn't run faster when
you upgrade a disk from 300MB to 1GB.
> > The NSF should cancel the 45 Mbps upgrade...
> Capacity is only one reason for going to T3; there are scientific
> applications that need 45 mb/s and more.
And I "need" a new car this year, too. What does it cost to go to
45MB? (The NSF contract with ANS must be a matter of public record!)
What other science goes begging because, rather than laying in a
satellite link between two sites who need 45Mbps, e.g. using existing
NASA science oriented network facilities, NSF funds an unwarranted
nationwide expansion?
