🚀 go-pugleaf

In article <1ejbadINNl57@early-bird.think.com>, sid@think.com (Sid Stuart) writes:
> 
> We are currently considering moving our backbone network from
> Ethernet to FDDI. As part of this change, we would also move
> additional computers onto the backbone. The resulting network
> would consist of three or four SGI's, 5-10 CM5's and 50 Sparc
> based computers: about 30 Sun 4/65's and 4/75's acting as
> fileservers and about 30 Sun 10/30's, 10/40's and 4/75's
> acting as compute servers.  The network would span two
> buildings and be about 1500-2000 meters in length. I expect
> the network to see 20-30 mbits of traffic during the day. 
> 
> Has anyone implemented such a network? What kind of problems have
> you seen? What are the performance characteristics? What kind of
> latency can I expect to see between stations under a load of 20-30
> mbits? Am I trying to put too much equipment on too large of a
> network or can FDDI deal with this distance/load?


The standard says you can have 2 kilometers between stations.  You can
probably cheat to get above 2000 m to deal with your buildings 2000 m
separation.

Are you going to put all 70 computers directly on the FDDI dual-ring?
With no concetrators?

If so, I would expect the ring to have a ring latency around 85
microseconds when the ring is THRU and around twice that when WRAPped.
Assuming all of the stations are doing ordinary TCP, UDP, and NFS kinds
of things, the average token latency of a ring at 30Mbit/sec or 30%
load and 100 usec latency is around 150 microseconds.   (The idea is
that in a 1 second, the ring will spend 300 milliseconds moving data
and the rest of the time spinning the token.  So solve for the number
of times the token will spin during the 700 milliseconds of idle time.)

At 30% load with ordinary stuff, any station will be able to transmit
any time a token comes around.  Thus, the network latency attributable
to FDDI in this situation averages 150 microseconds.

If you use concentrators, add around 1 microsecond per port on each
concentrator, whether active or not, to the ring latency, and recompute
the token latency.

You can get a lot more than 30 Mbit/sec TCP/IP/FDDI out of such a ring,
depending on the stations.

I know of FDDI networks like the one described, although they tend to
have more SGI, DEC or IBM boxes and fewer Sun boxes.

(If "20-30 mbits of traffic during the day" means 30 Mbit/day, then for
heaven's sake do NOT waste money on FDDI.  Use ethernets or T1 or even
56K lines.)


Vernon Schryver,  vjs@sgi.com