At 04:57 AM 6/18/00 +0300, Michael Black wrote: >On Sat, 17 Jun 2000, Christoph Lameter wrote: > >> Briding requires that the ethernet boards operate in promiscuous mode. >> The CPU has to inspect all traffic to check if it needs to be forwarded to > ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ >> another interface... Having Ethernet boards not needing much processing > > I know it's a big cpu time consumer. And I could switch to 3Com >EtherLink III (ISA) but I don't think it'll help a lot. I'll try in the >morning just with PCI boards (as Alan sugested), but I don't think it'll >help much. the traffic is still there. :) It feels like the processor it's >not cathing up with the traffic (4 any interfaces in use). btw when all >copying stops, everything's working great. > >> and a speedy CPU could be very helpful. But then there is the PCI > ^^^^^^^ >> bandwidth that will be another limiter. A bus mastering PCI card will not use CPU resources to move the data...a 16bit ISA bus is probably less that 25Mb/s best case with todays machines (its a secondary bus also), and with ISA the cpu is consumed doing the transfers. Since on a bridge you have to transfer the data twice, once on rx and once on tx, you can use almost all of your CPU just doing transfers. An 8 bit adapter is virtually unusable > > ah... how speedy? I think a 450mhz is quite a runner. and the 100mhz fsb >improves a lot. We have very busy ISPs bridging/switching 5 100Mb/s lans AND doing bandwidth management without the machine breathing very hard With PCI adapters its more of a pps issue until you get up to 400Mb/s on a single bus. db ------------------------------------------------------------------------- http://www.etinc.com T1/T3 boards for FreeBSD and Linux Multiport T1/T3 Routers Full-Featured Bandwidth Manager For Linux and FreeBSD - : send the line "unsubscribe linux-net" in the body of a message to majordomo@vger.kernel.org
