On Wed, 2003-11-19 at 21:42, Simon Allard wrote: > I have replaced NIC's as I thought it might of been the drives also. I > moved to the eepro100 cards. Same problem.
You should be using NICs with a poll-based driver, as opposed to an interrupt-based driver. This will preempt the kernel less often, and allow it to service the NIC more efficiently. The e1000 driver is excellent in this respect. We run more than 100Mb through a Linux router with a full eBGP table (~127k FIB entries) with no appreciable CPU consumption. The only time the box is substantially taxed is when a BGP peer flaps, in which case zebra consumes a lot of CPU power reconfiguring the FIB. It's a shame that the Linux kernel doesn't make the FIB structures accessible directly via an interface similar to /dev/kmem so zebra could simply mmap(2) it in and make large writes instead of small ioctl(2) calls. -- Jeff S Wheeler <[EMAIL PROTECTED]> Five Elements, Inc.
