janick@crk56.bnr.ca (Janick Bergeron 1617964) (12/14/89)
Looks like I am the de facto news administrator for a network
of (currently) about 200 Apollos... Now, I want to set-up things right.
To the best of my knowledge, here is how we're set-up
(No criticism about the set-up please):
o The mix is about 60% 4500, 30% 4000 and 10% 3000, all running SR10.1
o 99% of the WS are disked with 300M disks.
Most of them are 70-80% full. A few have 150M free.
o We have 1 4-CPU DN10000 with a lot of disk but it is used as
CPU server for heavy batch jobs.
Disk usage on that node varies a lot.
o WS are linked via Ethernet, on hubs of up to 64.
The hubs are hooked to a building-wide backbone.
Gateways connect backbones of several buildings together.
o The main feed comes from a off-site Sun via NNTP
Now for the wish list (in order of decreasing importance):
- Minimize disk space used for keeping articles
- Minimize network traffic when reading and transfering news
(main USENET feed nightly, internal stuff every 5 mins).
- Minimize response time when reading news
- Minimize the load on the server machine
Here are a few observations made on a temporary set-up:
o rrn/Pnews are VERY slow to start on client nodes. But once it's
started, everything goes smoothly.
(I think someone tracked that problem to the gethostbyname()
call in the NNTP stuff...)
o My clients seems to prefer to go into VT100 mode, crp on the server
then use local rn. One would think that that would minimize traffic
as only screen update needs to be transfered, not entire articles
(especially if you use '=' a lot)
o My node is the server and I have 3 regular NNTP readers and 2 readers
who crp onto my node and use local rn. I don't see a significant
decrease in performances.
o I am currently receiving news from a main feed and in turn feeding
two other servers. Transfering news does significantly impact on
the performance of my node.
o I was fairly successfull at managing 4 servers from my own server
which is the USENET gateway in/out of the building. Using a
building-only distribution, tasks like newsgroup creation and removal,
article "unjunking" and auditing was not too painful.
Given these constraints and objectives:
+ What should be a good ratio of server/clients ?
+ Should we crp onto the server nodes instead ?
+ To how many other server should a server transfer news ?
I suppose that providing a newsgroup on demand rather than a full
feed (regardless whether a newsgroup is read or not) would greatly
reduce traffic; but that would give me a headache just managing the
sys files!!.
Thanks for any help. Please reply via email to address below.
Follow-ups (if any) directed to news.admin only.
--
Janick Bergeron Bell-Northern Research, Ltd Ph.: (613) 763-5457
VHDL Tools P.O. Box 3511, Station C Fax: (613) 763-2661
Ottawa, Ontario, Canada, K1Y 4H7
janick@bnr.ca library disclaimer; use disclaimer.all;