IRIS clusters

[Mike Gigante]

CMSDS at UGA.CC.UGA.EDU (David Stewart) writes:


>    I am interested in finding out about "clustered" IRISes, that is
>a group of IRISes within one logical group (lab).  It is possible that
>we may be able to purchase a number of Personal Irises to go along with
>our existing 4D/70GT.  My question is:  Should we go with disks or
>diskless (they will be linked via FDDI).  What experiences have others
>had with similar setups (even those not linked by FDDI).

Well, we have 20 SGI machines in a single lab at the moment. It will
split into a group of 10, another of 6 and the rest sprinkled around
sometime next year.

The server is a 220 (SMD), most of the PIs are currently diuskless but I
have abandoned the diskless setup and have already ordered 200Mb system 
disks for most of them.

In my opinion, a diskless PI is a bad idea. With the current setup
(NeWS, the share tree, *symbolic* links fromn the client to share), the 
net traffic from diskless to server is just too high. (I put a sniffer
on the net and watched the traffic). Login from a diskless is ridiculously
slow (this should change once SGI go over to X right?). Paging across the net
is also pretty slow. 

This is particularly bad if you have 8Mb PIs (I have upgraded all PIs to 16Mb)
For example, if you have NeWS, Xsgi, 3 wsh, clock, you aready need to page!!!!!

(SGI: please compile wsh with the shared library if it isn't now!!! It is 
ridiculously large for what it does...)

This is not a flame, merely a comment based on 15 months experience with
diskless. The incremental cost of a 200Mb system disk is *REALLY* worth it.
I would like to emphasise the point *very* strongly.

To be fair, a 16Mb PI with 3.3 is a much more usable diskless machine than
the 8Mb 3.2 machine. However, now that disk prices have come down, the balance
is definately in diskfull's favour.

Mike Gigante
RMIT Advanced Computer Graphics Centre