stevel@eleazar.dartmouth.edu (Steve Ligett) (09/26/89)
In article <1462@intercon.com> amanda@intercon.com (Amanda Walker) writes: ... >Aside from that, though, the only benefit I can see of the Asante board >over anything else is that uses a 32-bit data bus to its buffer memory, >instead of the 8-bit wide one that most other boards use, which could >give a minor speed improvement in some applications. It doesn't seem to >matter much in normal use, though. > >-- >Amanda Walker >amanda@intercon.com For more insight into this, here's a note I got from our system architect: Dartmouth uses Mac-IIs for IP and AT routers (function similar to Apple's internet router for the II), with up to 6 enet cards. The bottleneck in this application is enet card throughput, so we've done a lot of careful analysis of the various vendor's cards/drivers. More specifically, for our router the bottleneck is the data rate between the card's RAM and system RAM, over the nubus. As you probably know, there is no DMA on the Mac-II so the CPU must copy data to and from the enet card's local RAM. The Apple and Kinetics cards do nubus transactions in 16-bit units, even though the bus is 32 bits wide. In addition, the Kinetics card (at least the one with the Intel chipset) has some bugs when exposed to very heavy load, ie >200 pkts/sec. It can hang and drop more pkts than it should. I've never tried to break the Apple card, so can't say if it has bugs too. The Asante card does *32* bit xfers across the nubus, so for *OUR* application its a big win. Its very hard to get the Asante card to drop back-to-back packets or hang, at least with our application (which is very quick about moving incoming data into system ram, thus freeing up card buffer space). Finally, the Asante is the least expensive. We've been paying about $352 for them. | Apple and Kinetics | Asante pkt size | pkts/sec bytes/sec | pkts/sec bytes/sec ---------+----------+----------+----------+--------- 39 645 20k 1080 40k 529 290 150k 700 360k 1040 258 264k 534 545k The above figures are quite accurate, having been measured with our Network General Sniffer and verified by the router's own internal instrumentation. They are sustainable *THROUGHPUT*, ie a pkt or byte read on one card, copied to another card, and then written is counted just once. A normal application running as an endpoint in a network connection should see almost twice as much bandwidth (unverified). Don't worry that 645 pkts/sec doesn't quite add up to 20kb/sec... headers etc are counted in one case but not the other,etc. The big caveat about these results is that I'll bet that most 'normal' applications are not nearly so sensitive to nubus throughput as our router. After all, a router doesn't process the data at all... all it does is read and write it. With more processing per byte, the advantage of the Asante card will be diminished. One guy at CMU told me he didn't see *ANY* improvement with the Asante. He was CPU bound, it turned out... Philip Koch Philip.Koch@Dartmouth.EDU -- Steve Ligett steve.ligett@dartmouth.edu or (decvax harvard linus true)!dartvax!steve.ligett