Severe packet loss on fxp interfaces with the new bridging code...

Matthew Dillon dillon at apollo.backplane.com
Sat Dec 24 14:54:51 PST 2005


:It looks like this is the problem after all.   My machine has 128M of
:ram, which results in 1504 mbuf clusters by default.  In operation,
:I have never seen it break 1000, so I assumed that this was not
:the issue.  Since the magazine size for clusters ends up being 256
:though, it appears that there is not enough wiggle room for the
:objcache here.
:
:After increasing nmbclusters though, things work perfectly.  Still,
:it might be useful for this knob to take into account the number
:of cpus and magazine size.
:
:Anyways, enjoy the holidays!
:
:Chris

    Ok.  That makes a lot of sense.  The granularity is too high and
    the objcache can't maintain enough magazines on each cpu.  We have
    a minimum calculation for that in the kernel which I'll have to
    bump up, but I'll do it after the release.

    Do you still want a leaf account?

					-Matt
					Matthew Dillon 
					<dillon at xxxxxxxxxxxxx>





More information about the Kernel mailing list