Am 22.08.2010 um 23:34 schrieb grarpamp: >> FreeBSD does that differently. Afaik actually only make sense when you go >> tickless with the Kernel which FreeBSD hasn't done. So their way of reducing >> networking load is by polling the card instead of having it send thousands >> of interrupts. It works with igb and ixgb, I'm not sure about e1000. > > Yes, they implemented polling for many NIC's. And toggleable on the > fly with ifconfig. Hz is adjustable if need be. There are also kernel, > loader and sysctl bits. And things like interrupt coalescing and > loadable microcode. All depends on the > NIC... Check the man pages, driver source/readme's and kernel config. Yes, but remember, that setting kern.hz affects all cores. What probably also helps a lot with polling is probably increasing buffer size to 32 or even 64 megs. I just did double check on the manpaes for ixgb, em and igb. All of those drivers support polling. The devices mentioned are supported by if_em and polling is supported. > >> the current FreeBSD Stable Release doesn't work with Intel NICs _again_ > > I can't confirm. Given that Intel is one of the top three or so most > popular and performant nics, I highly doubt RELENG_8_1_0_RELEASE > shipped unable to pass packets. So the FreeBSD team should postpone a code freeze just becaue Jack Vogel isn't abled to follow the committing rules everyone else follows? After all Intel will earn a bad reputation for shitty drivers, not necessarily FreeBSD. That said it sadly iwill freeze with releng_8_1_release. See the mailing lists for reference. Googling "Jack, did you break em(4) (or lem in this case) again? :-)" is probably the best starting point for it actually. > > Either way, if you cvsup or cvs co the driver to RELENG_8, you > will get a commit as of 10 days ago, as opposed to the last one 4 months ago. > And you might still be able to drop in a version from Intel's site too. I just checked, there's an errata since a few days that fixes it in the 8_1 branch. Backporting from 8-stable also works, pulling the source from intel, too but imo thats not the way of handling things on production systems.