CY13@TE.CC.CMU.EDU.UUCP (02/26/87)
Here at last is my promised summary of the replies to the
DELNI Bottleneck problem...it appears to be nonexistent. Several
people commented on how the situation is set up, which was not
necessary. Some talked about the Heartbeat/Repeater/Delni problems,
which are not included. I thank all of you who replied. The
identities are edited out by request.
Curt Yeske
Technical Administrator
Carnegie Mellon Computing Center
Disclaimer: The opinions expressed are my own and are not Carnegie Mellon's.
The facts are figments of my imagination anyway.
*** My Message: ***
I realize that this question may not be exactly right for this
group, but I think that there may have been discussion on this issue in the
past -- and I need to be updated.
Currently we run a many workstation (300+?) and central fileservers
(14) on a single ethernet (right now we can't change this basic configuration).
My question revolves around the fileserver connection to the network.
Currently we have 14 drop cables going into 2 DELNI's. The question I have
is if I tap the fileservers directly onto the ethernet, not using the DELNI's
would I gain in performance? Do DELNI's get overloaded? Do they cause
increases in network overhead? Any help in this direction would be most
helpful, especially if you have had large amounts of data passing through
a DELNI.
Please send mail to cy13@te.cc.cmu.edu, and I will post a summary.
Thank you very much,
Curt Yeske
-------
*** The Replies ***
delni's provide identical services to trancievers. it is
inherent in the definition of trancievers that they can't run
"slower". I suppose there will be a few nanoseconds propagation delay
added by the extra chips, but this should be less than that added by
the 16 meters of wire that you would have to have between the
tranceviers.
again the major thing to watch with delni's (and ther kind: see below)
is that the tranciever cable length's add. thus there are limits to
the amount of cascading that you can do with them (DEC spec's prohibit
delni -> delni -> coax connections. as was mentioned before, such
things work in small nets, but when things get bigger, watch out.)
btw: there is an english firm making a similar box (they call it a
fan-out) it is distributed in this country by Cabletron (in Nashua
NH). same number of slots, a light to show when there is traffic on
the net (fun to watch) and about $200 cheaper than DEC. They also sell
trancievers with set of diagnostic (tx rx coll) lights on them (which
helped me find the machine that was occaisonly jamming the net.), a
TDR (time domain reflectometer) that does not require you to power
down everything connected to the net before you can check things, and
a tranciever tester. (as always I have no coneections with cabletron)
<dp>
*** ***
There is no throughput limitation in a DELNI. The only way that DELNIs
can reduce network throughput is if the installation violates the rules,
in such a way that collisions increase. To fulfill the rules, the main
Ethernet cables must be 1 500m run, plus any number of additional 500m
runs that attach to the root cable through repeaters. Then, to this
system of cabling, any number of end nodes and DELNIs may be attached
anywhere. The only thing that may be connected to a port on a DELNI is
an end node. Also, end nodes that attach to DELNIs rather than to the
trunks themselves should have tranceiver cables that are 15m shorter
than the tranceiver cable length normally allowed by the manufacturer
of the Ethernet interface in the end node. Since the max. allowed
tranceiver cable is 50m or less (depends on the interface board that
you use), a station connected through a DELNI should use 35m or less.
All of these numbers are off the top of my head. If you need to know,
look them up in the relevant documents. Also, an Ethernet installation
that violates the spec., but in such a way that excess collisions do not
occur, will work just fine! Certain tranceiver cables can be 500m if
the main trunks span less than 1000m, for instance.
Check your Ethernet statistics -- collisions, "late collisions", and
average loading. "Late collisions" indicate an illegal cabling arrangement
or a flaky interface. (You can disregard very small error counts,
though.) If you see abnormally high collisions and high average loading,
you may want to partition the network. Put half the workstations on an
Ethernet, together with the fileservers that serve them. The other
stations and fileservers go on another Ethernet. Then, attach the two
Ethernet trunks with a DEC LANbridge 100 (accept no substitutes). This
is the only full-bandwidth Ethernet bridge that I know of. Even if you
blew it in the partitioning, you don't loose throughput, because the
bridge is not a bottleneck. But you won't gain anything unless the
network traffic can be divided, so that most traffic on Ethernet A doesn't
get copied to Ethernet B.
Hope this helps,
*** (an addition) ***
Subject: Oops...
I believe I made an error in the statement about throughput through a
LAN Bridge (which is a store-and-forward device, unlike DELNI).
When you split a heavily loaded network with a bridge, the traffic through
the bridge is subject to the same arbitration onto the target Ethernet
cable as a regular end station. So, although the bridge hardware can
forward packets between cables at the full Ethernet rate (even for minimum
sized packets), the throughput may be limited by arbitration to put
each packet onto the target Ethernet trunk. If the target trunk is very
busy, then at times, the bridge can't pass the full traffic of the
originating trunk onto the target trunk.
In short, the sum of cross-bridge traffic onto a trunk and all of the
traffic that originates on a trunk is limited to 1 Ethernet's worth of
bandwidth. You only get higher aggregate bandwidth on both trunks
combined when not all of the traffic has to cross the bridge.
I know this is noise, but I don't want to misrepresent the facts.
*** ***
Subject: An alternative.
Curt,
The DELNI performs the same functions of eight Ethernet transceivers. There's
no real additional overhead. To improve performance at a site such as yours,
run dual ethernets to isolate LAT traffic from cpu-cpu traffic.
************************* (data traffic ethernet)
$< $< *
+-----+ +-----+ +-----+
|VAX_0| |VAX_1| |W/S | 'workstation'
+-----+ +-----+ +-----+
$> $>
************* (terminal traffic ethernet)
/ | \
/ | \
/ | \ '$<' circuit cost is lower than
VT0 VT1 VT2... '$>' this circuit. Forces cpu to
use the data traffic ethernet for
\|/ initiating logical links.
*** ***
Curt,
I can think of no way in which a DELNI can introduce a performance bottleneck
on an Ethernet. Transceivers maintain no state and since only one packet
can be on the net at a time anyway, there can be no issue of datapath
bottlenecks.
*** ***
No, there is no gain in performance. DELNIs are essentially passive,
and cannot be overloaded - they have the same performance as an
Ethernet, or perhaps a little bit better because of less propagation
delay or less noise.
Anyhow, I am impressed by the big number of things you run on your
Ethernet... Seems you'll have to install two Ethernets ? I would be
interested to know how you solved that problem...
*** ***
-------