At 05:13 PM 6/18/00 -0400, Donald Becker wrote: >On Sun, 18 Jun 2000, Andrew Morton wrote: > >> Alan Cox wrote: >> > >> > You want to get off ISA, even to low end PCI bus masters (just avoid NE2K PCI >> > clones) >> >> I have a 400MHz Celeron UP here which is running at 100% when sending 10 >> MBits/sec with an EISA card (cs89x0). > >Unless the cs89x0 driver is doing something unusual, or you application is >CPU hungry, that sounds a little off. We could saturate a 10Mbps Ethernet >with a 386. > >> With a PCI NIC (3c905) it will do 100Mbits/sec at 60% CPU. > >Also high: you should be able to drive at least two, almost three, 100Mbps >channels with a P5-100. one issue, that Allan shrugged off the other day, is that linux insists on queuing everything at the MAC layer, even known local traffic, so there is substantial undue overhead when bridging. You are not just "moving" 60Mb/s, you are queuing all of your traffic that should be filtered by properly implemented code. We can bridge over 100K packets per second on a 500 mhz system at about 80% utilization with eepro100s on BSD systems so the physical capabilities are there. We rate a 600Mhz machine to be able to bridge 150Mb/s with average size packets and still be usable for other activities. Linux achieves similar numbers when the eepro driver doesnt lock up :-) Dennis - : send the line "unsubscribe linux-net" in the body of a message to majordomo@vger.kernel.org
