Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: version B 2.10.2 9/18/84; site Glacier.ARPA
Path: utzoo!watmath!clyde!burl!ulysses!allegra!oliveb!Glacier!reid
From: reid@Glacier.ARPA (Brian Reid)
Newsgroups: net.news
Subject: Re: netnews eats 750s for lunch?
Message-ID: <11798@Glacier.ARPA>
Date: Fri, 13-Sep-85 23:40:59 EDT
Article-I.D.: Glacier.11798
Posted: Fri Sep 13 23:40:59 1985
Date-Received: Sun, 15-Sep-85 00:22:59 EDT
References: <1996@amdahl.UUCP> <791@vortex.UUCP> <626@bu-cs.UUCP> <1090@ulysses.UUCP> <116@l5.uucp>
Reply-To: reid@Glacier.UUCP (Brian Reid)
Organization: Stanford University, Computer Systems Lab
Lines: 36
Summary: 

Glacier is a 750. Until today we had 2Mbytes (today we got a new backplane
and we have 4Mbytes).  We feed netnews to 12.5 sites (12 full feeds and one
partial feed, plus about 20 local SF Bay feeds).

Glacier is also the primary computing resource for about 100 people; during
the day there are typically 30 jobs logged on, and at night there are
perhaps 5 to 10 jobs logged on.

The load factor on Glacier can get pretty grim at 3 in the afternoon, but
netnews is never the culprit. Well, perhaps people reading news with "rn" (a
resource hog) are part of the problem, but news transmission is not a
problem.

What I have found is that distribution of netnews over local area networks
is essentially free on a machine like a VAX that has a high-performance
asynchronous parallel data path to the network. The amount of CPU spent is
almost unmeasurable; Glacier spends perhaps a total of 15 minutes a day,
elapsed time, feeding these 7 LAN sites.

Telephone line feeds, however, are very expensive. It burns a lot of CPU
and causes a lot of context swaps. All of you know this, though.
We don't do any phone line feeds during the day (except ba.seminars and
things like that). If we tried to do phone line feeds during the day it
would make a significant dent in our machine.

The moral? Technology, such as our 3 and 10 megabit Ethernets and decent
system software, can virtually eliminate the compute cost of transmitting
news. There is still a cost associated with receiving news, caused by all of
the forks and execs in recnews rnews, inews, unbatchnews, cunbatch, and all
of those programs, but each site only has to do that once.

I don't think 2400-baud phone lines would cure anything. Some hardware that
enabled block synchronous transmisison over them would help a lot.
-- 
	Brian Reid	decwrl!glacier!reid
	Stanford	reid@SU-Glacier.ARPA