vince@bcsaic.UUCP (Vince Skahan) (08/01/89)
We run a variety of rings of Apollos and a number of ethernet-only
Apollos on a campus ethernet. Each Apollo can talk to each other via
Apollo token ring protocol via the appropriate rtsvc commands. We have
experienced what we think are performance problems between rings
(relative to the speed of a token-ring) that appear to be far in excess
of the expected difference (10MB vs 12 MB coax).
Our ethernet connections are either straight drop cables off of the
campus fiber, thinwire ethernet ultimately off of a tap into the fiber,
or LatticeNet over twisted pair.
Apollo informs us that running Apollo over Latticenet is not supported
since that means it's over twisted pair (and that's what's really not
officially supported).
Does anyone have any information about any of the following ?
- should we see definite slowness between rings regardless of
how the rings are connected to ethernet ? we remember a definite
perceived slowness of "native apollo" relative to the old "etherbridge"
product.
- should we see a special drop in performance over Latticenet
(is anyone using Apollo over laticcenet. ?)?
- if Latticenet is "real ethernet" why isn't it officially
supported? We need good technical reasons to make the networking folks
take notice.
- should we see any difference in performance from node to node
if the nodes are on Latticenet relative to them being on drop cables
- should we see perceptable (real perceptable) slowness going
over ethern to copy files, do backups, copy large trees with lots of
small files, etc. ?
- what is the REAL minimum speed of the ethernet to enable
reasonable apollo-apollo communications without experiencing network
timeouts ? We have one bulding remotely located with a link that is
about 1/2 of T-1 speed. We experience lots of network timeouts between
down there and the rest of the ethernet. When you consider that we have
a registry site down there, it takes forever to add accounts, change
passwords, etc.
===> any help would be greatlyapreciated. Feel free to either e-mail or
post followups as you wish...
--
Vince Skahan Boeing Computer Services - Phila.
(215) 591-4116 ARPA: skahan@boeing.com
UUCP: bcsaic!skahan
Note: the opinions expressed above are mine and have no connection to Boeing...krowitz@RICHTER.MIT.EDU (David Krowitz) (08/02/89)
A few comments ... 1) MIT's telecommunications office runs the campus-wide ethernet system which our Apollo ringnet is attached to via a DN660 gateway. The telecom office states that the maximum throughput of an ethernet segment is 3 MB/sec on a 10 Mhz thick ethernet cable. This is due to packet collisions eating up network bandwith. Thus, the maximum capacity of our building ethernet (a thick cable network) is 1/4 of the capacity of our 12 Mhz ringnet. (please excuse my typo ... ethernet throughput is 3 Mhz/sec). Apollo talking to each other over native ethernet connections will see a lot of collisions on a heavily loaded network. Apollos on native ringnets which are attached to each other via an ethernet backbone will also see a bottleneck. 2) Network throughput is directly proportional to the number of gateways seperating the nodes. Since the gateways are Apollo workstations and not dedicated hardwired boxes, the time lag to pass a packet from the ringnet interface to the ethernet interface is considerable. This is why diskless node service, /com/lcnode, /com/lusr, and other time dependent network services only work on your local network even if internet routing has been enabled. 3) Paging from diskless nodes and remote file access can eat up network bandwith in a hurry. Try running /com/dspst -a sometime while your node is copying a file to another node's disk. The number of paging requests per second can easily exceed 100/sec. If we assume a page size of 1024 bytes, 8 bits/byte, and 100 pages/sec we see that a single node can consume 0.8 Mhz of bandwith -- almost 25% of the usable bandwith of an ethernet! Now try copying two or three files at once (in seperate windows). You can easily get a single node to generate 200 or 300 paging requests per second!! -- David Krowitz krowitz@richter.mit.edu (18.83.0.109) krowitz%richter@eddie.mit.edu krowitz%richter@athena.mit.edu krowitz%richter.mit.edu@mitvma.bitnet (in order of decreasing preference)
lori@hacgate.scg.hac.com (Lori Barfield) (08/03/89)
Well, here in my group this discussion went around awhile ago. We have an internet of two rings, and the communication between the two was at times 100 times slower than normal inter-ring talk. One engineer ran tests on our DN3000 on SR9.7; he claims that the Suns on the *same* Ethernet are much, much faster than the Apollo is. When I sent him a copy of David's post, this is what he replied: >In article <8908012221.AA05657@richter.mit.edu> krowitz@RICHTER.MIT.EDU (David Krowitz) writes: > >[...]telecom office states that the maximum throughput of an ethernet > > segment is 3 MB/sec on a 10 Mhz thick ethernet cable. This is due > > to packet collisions eating up network bandwith. Thus, the maximum > >[...] >On our network we typically run at less than %10 of ethernet capacity >according to our network monitoring tools. Collissions are VERY rare >events, I estimate I see a collision once every 10 minutes. You can >see this by bringing up the performence meter and selecting the >collisions or by looking at one of the ethernet transceivers with >the LED lights on it. > >Also I used to run 10-12 diskless suns on an ethernet, doing all >their swapping and paging over ethernet and never used more than >%20 of ethernet capacity. I suspect the Apollo interface to >ethernet is just plain Lame. So the problem apparently isn't collision. Any comments? Is SR10 any better for Native? ...lori
tim@tim.UUCP (Tim Giebelhaus) (08/05/89)
In article <13604@bcsaic.UUCP> vince@bcsaic.UUCP (Vince Skahan) writes: >Does anyone have any information about any of the following ? > - should we see definite slowness between rings regardless of >how the rings are connected to ethernet ? we remember a definite >perceived slowness of "native apollo" relative to the old "etherbridge" >product. Of course any gateway is going to slow down the data. If the gateway is busy doing something else, or is asked to gateway more data than it can handle, the tcp, routed, and other supporting processes will not get the data through very quickly. > - should we see a special drop in performance over Latticenet >(is anyone using Apollo over laticcenet. ?)? I'm not sure what Latticenet is. If it does not work at a speed of 10 mb/s they you are very likely to have problems. The time outs and such on Apollos are written with the assumption that you are running raw data of 10 mb/s. > - if Latticenet is "real ethernet" why isn't it officially >supported? We need good technical reasons to make the networking folks >take notice. If Latticenet is "real ethernet" and follows the 802.3 specs to the letter, then it should not be a problem. > - should we see any difference in performance from node to node >if the nodes are on Latticenet relative to them being on drop cables Only if Latticnet is not reliably forwarding the packets or is not forwarding the packets as fast as "real eithernet". > - should we see perceptable (real perceptable) slowness going >over ethern to copy files, do backups, copy large trees with lots of >small files, etc. ? You point out small file in particular. Of course the opening of lots of files is going to take longer than opening one file and then copying the same amount of data. Also, I sure you understand heavy use such as backups is bound to slow up the network and the node. > - what is the REAL minimum speed of the ethernet to enable >reasonable apollo-apollo communications without experiencing network >timeouts ? We have one bulding remotely located with a link that is >about 1/2 of T-1 speed. We experience lots of network timeouts between >down there and the rest of the ethernet. When you consider that we have >a registry site down there, it takes forever to add accounts, change >passwords, etc. Apollo states in it's manuals that you must use at least a T-1 phone line if you are using phone lines. I know of sites which use 56kb, but this is not supported and problems are seen with it. I would recommend a less robust protocol for slower speeds such as NFS. If you have 9.7 registries then you indeed will take a while to do registry operations as, up till sr10, the entire registry was copied over and then copied back. I would recommend the reading of all the SR10 networking manuals. Especially the planning manual may be useful. -- UUCP: uunet!hi-csc!apcimsp!tim ARPA: tim@apollo.com Contents of this message has nothing to do with work.