orand@kuhub.cc.ukans.edu (03/30/90)
I need some expert advice on networks. We have a lab with 16 DN 4000's networked with Ethernet. We are having severe problems with network speed and are contemplating going to Apollo Token Ring. I for one, think this is an excellent idea but would like to get some "real" opinions. The question is: "Ethernet or Apollo Token Ring?" "Why?" Brady... =========================================================================== Brady Orand - University of Kansas Computer Center Lawrence, Ks. 66045 ORAND@kuhub.cc.ukans.edu Work: (913) 864-0490 Home: (913) 749-1341 ===========================================================================
wjw@eb.ele.tue.nl (Willem Jan Withagen) (04/09/90)
In article <21091@joshua.athertn.Atherton.COM> joshua@Atherton.COM (Flame Bait) writes: >chen@digital.sps.mot.com (Jinfu Chen) writes: >> Also in ethernet environment you can't never get over 50% of the >> spec'ed speed due to the collsion. > >Wrong answer. The right answer is that the maximum real world throughput >is at least 8.9 Mbits over the wire and over 1Mbyte task-to-task, or over >85% of spec. Interesting details: > 2 processes talking to each other. > Two 20MHz 68020 machines > Running 4BSD UNIX (I assume 4.3). > The test was done in Sept. 88 > Using an AMD LANCE chip with a receive buffer of 10 packets or more. > I wonder wether this is the right correction. One of the key remarks in this is 'collision'. Now collisions do not happen on an ethernet with only two communicating processes. As soon as more processes want to communicate at the same time, their packages start to collide and are resent ( even more than once ). And this fact is specifically ignored in the above test. Hence, the 8.9Mbits spec tells you taht an 10Mbit ethernet without collisions does not even get close to an 10Mbit performance. Using Jinfu Chen estimate gives this ~4.5Mbits real troughput on a wire with more than 2 communicating processes. Well at least this is the story, as I tell it during my classes. If things have altered this significantly then I must have been sleeping. And will now be awakend by the flames of the net. Greetings Willem Jan Withagen Eindhoven University of Technology DomainName: wjw@eb.ele.tue.nl BITNET: ELEBWJ@HEITUE5.BITNET room EH 10.10 Digital Systems Group P.O. 513 Tel: +31-40-473401 5600 MB Eindhoven The Netherlands
krowitz%richter@UMIX.CC.UMICH.EDU (David Krowitz) (04/09/90)
MIT telecommunications office estimates that in a *REAL* "real world" situation (ie. several dozen Sun-3's, DEC Vaxstations of various sorts, and IBM PC's running project Athena software) they never get more than 30% of the specified ethernet performance. I can believe that you'd get 85% of the spec with only two nodes talking -- you would not get vary many collisions -- but try putting half a dozen diskless Sun-3's talking to 4 or 5 file servers all doing geophysics, oceanography, or meteorology applications. It's a whole different ballgame. A single node reading a remote data file, or doing diskless paging, can easily generate 100 1kbyte packets per second. This is ~1/10th the capacity of an ethernet. If three nodes (out of a few dozen) were running like this, and if they were somehow avoiding any collisions (which would eat up extra net bandwidth), they would be consumming 30% of the network's capacity. Any addition node would have at least a 30% chance of causng a collision ON EVERY SINGLE TRANSMIT IT ATTEMPTS! Each collision with a 1Kb packet consumes 0.1% of the network capacity. At 100 packets/sec, this one additional node is eating up a minimum of 3% of the network capacity in collisions by itself. Add routing daemons, rlogin/telnet sessions (which use a TCP/IP packet for every single character you type!), and twenty to thirty nodes all running NFS with a 50-50 split between diskless nodes and severs and the network capacity eaten up by collisions is substaintial. -- David Krowitz krowitz@richter.mit.edu (18.83.0.109) krowitz%richter.mit.edu@eddie.mit.edu krowitz%richter.mit.edu@mitvma.bitnet (in order of decreasing preference)