jod (09/15/82)
We have Interlan Ethernet equipment in "production" (that is, we depend upon it) use on 11 VAX11/750s, 1 VAX11/780, an Apollo Domain, 3 PDP11s, and 2 DEC-2060s. For the last several years we've been running Chaosnet among a smaller set of hosts on hardware we built. We started conversion in May to Interlan hardware, retaining the Chaosnet protocols and extending them as per the Xerox/Symbolics/MIT Chaosnet "standard". We encountered a problem on the Q-bus board which Interlan found and has fixed. Since then we've had no problems whatever. Daily traffic on the net ranges from 500000 packets to about 12 times that. We use the net for the normal assortment of things, including heavy remote-printer and remote-magtape traffic. There are currently 2 disconnected subnets; a group of 5 VAX 750s are out of reach of the main cable, which connects hosts in 3 buildings. A point to point link (GE GEMlink) will be installed soon as a bridge. We noticed substantially lower host loads on VAXes than with our homebrew hardware (which resembled the 3Com board in architecture -- a shared-memory setup). We've also noticed dramatically lower retransmission rates -- due mostly, we believe, to the receiver buffers in the Interlan board. But we haven't seen much speedup on individual connections -- maybe 20% (to an average of ~180KB/s). Since we've never done protocol tuning or monitoring, this isn't surprising. The limit is NOT the hardware. Pat O'Donnell (pao@xx) has done some direct comparisons of the 3Com and Interlan controllers on a VAX11/750. I believe his results indicated that in transmit performance alone the boards were comparable in throughput; 3Com beat Interlan for short packets, vice versa for long packets, with breakeven at about 200 bytes. However, the Interlan board consumed 20-30% of the CPU (VAX750) while the 3Com consumed ~70% while running the test. Symbolics (the Lisp Machine people) offer Chaosnet supported for all the systems I mentioned except the Apollo, if you're interested.