dennis@PEANUTS.NOSC.MIL (Dennis Cottel) (02/24/89)
About a month ago I asked the network at large for comments on experience with Apollos connected directly to an Ethernet rather than the Apollo Domain Ring. The responses are included here. After that are some notes I made in talking to some local organizations who have some nodes in this configuration. I also heard a rumor that there was some kind of Apollo report on the subject, but I was unable to find any such document, either through the Apollo hotline or my local sales office. Dennis Cottel Naval Ocean Systems Center, San Diego, CA 92152 (619) 553-1645 dennis@nosc.MIL sdcsvax!noscvax!dennis ------------------------------------------------------------------------- >From: tolerant!procase!ajs@ucbvax.Berkeley.EDU (Andrew Jay Schneider) We're one of the first large (30 nodes, hah!) users of Etherneted Apollos. We went through some hard times with their defective Ethernet microcode driver but it is pretty stable now. Make sure you either run SR10.X or that you have a version of /sys/ethernet8_microcode from their patch tape, the stock 9.7 version will not work under any real load. Apollo claims that the degradation is something like 20% compared to the ring but my gut feeling is that it is really less. Maybe 5% to 10% for normal interactive use. We don't usually run diskless nodes here so I don't have any direct feel for that case. You should know that the 400 series along with the DSP-80's and 90's won't function on an ethernet along with native mode machines. If you have any of these machines they'll have to be on a ring with one of the native machines functioning as a gateway. We do find accesses through the gateway to be about 30% degraded. All in all I think going ethernet for us was the right choice. We are a mixed Sun, DEC, HP, Sony and Apollo shop. All of these machines coexist peacefully on our single network. We can move machines from office to office without breaking the ring. We can now monitor the network with comercially available tools. We use TCP/IP and NFS to connect the Apollos with the others. Also I've found that the reliability of the ring left something to be desired. At my old company we had 97 Apollos on one ring and about once a month something would bring the net down. In the almost two years we've been running ethernet we've had only two complete net outages of about 5 minutes each. (Both caused by a stray piece of braid during tranciever installation.) Oh, yeah, don't go with thin net. In my opinion, it is no more reliable than the ring. ------------------------------------------------------------------------- >From: mcvax!cernvax!achille@uunet.UU.NET (Achille Petrilli) Hi, we have some 60 nodes on a ring and 20 on ethernet. Transfer on ethernet is somewhat slower than on the ring, but I guess is due to the number of machines hooked up on eth rather than anything else (we have a total of ~ 500 hosts on eth running decnet, tcp/ip and apollo/domain). Tcp/ip is much better on eth than on ring. It seems also that Apollo protocol on eht is exactly the same as on the ring and that penalizes raw speed on eth as shown by rtchk vs. ping. We recommend everyone now to buy eth boards rather than ring, this makes life much easier for tcp/ip and makes troubleshooting easier as well. ------------------------------------------------------------------------- >From: ames!rtech!squid!edg@ucsd.edu (Ed Goldman) I don't know much about a direct comparison between ethernet and ringnet thruput where there would only be apollo's on the ethernet, but ... I do know from taking apollo's comm class that ring-net packets are encapsul- ated in ethernet packets. If you have other tcp-ip hosts on the ethernet they'll see the ring-net encapsulated stuff as errors which could hurt the overall thruput on the ether. I image there's at least a small performance hit in that the apollo's need to de-encapsulate the ethernet packs to get at the native ring-net pack. ------------------------------------------------------------------------- >From: giebelhaus%hi-csc.uucp@uunet.uu.net (Timothy R. Giebelhaus) I'm glad you asked. I often help people who over load their network. It is much harder to rewire a building than to wire it the first time. A token ring network can handle a heavy load much better than an ethernet (references available on request: that is, they're at work as I use them pretty often). The fact that a token ring can handle a heavy load is a general fact; not specific to Apollos. Most networks of workstations generate a pretty heavy load. Diskless nodes will generate traffic very quickly. Centeralized CAD libraries can also generate a lot of traffic. The newer multi-node backup software will generate much traffic. So, unless your nodes pretty much keep to themselves, I would use two ethernets or one ring. You could probably get away with one ethernet, but you will quite likely notice degradation. If you have very heavy traffic now, you will want three or four ethernets or one ring. There is no doubt that one network is easier to administer than several gatewayed networks, so I would use the ring. ------------------------------------------------------------------------- Notes on phone conversations with local San Diego companies on the subject of running Apollos on Ethernet: Theresa Macklin, General Dynamics, San Diego, 573-3612, Jan 25, 1989 They have assorted machines on a collection of rings and Ethernet pieces using Ethernet on twisted pair. There is a backbone ring of servers. Need to use the NS_HELPER. They feel it is slow. How slow? They did some "preliminary" benchmarking comparing performance to Apollo ring network: 1/3 as fast through two gateways 1/2 to 2/3 as fast between two nodes on the same Ethernet depending on the Ethernet traffic load A diskless node to a host on the same piece of Ethernet is ok. If set up right, it is acceptable. It is much easier to add or move nodes. John Reddan, Syscon, San Diego, 296-0085, Jan 26, 1989 6 nodes (4 diskless) and 2 Suns on Ethernet. They've had problems with NFS to the Suns. Never run on an Apollo ring so can't compare. Works with no problems. Mike Markley, UCSD, San Diego, 534-2221, Jan 26, 1989 They have only a couple machines on Ethernet. Found it really slow, but they're on the campus Ethernet which is probably heavily loaded. -------------------------------------------------------------------------