On Fri, Jun 22, 2012 at 04:58, Brad Smith wrote: > On Thu, Mar 01, 2012 at 03:28:13PM +0100, Mike Belopuhov wrote: >> This is a well-known from thib and dlg originally with a length fix >> from yours truly, that marginally doubles througput (from 300kpps to >> 500-600kpps on selected hardware). >> >> The idea is to save an IP header and 8 bytes of payload (good enough >> for tcp state tracking) instead of recommended 68 bytes. >> >> We've been running with this diff for quite some time and haven't seen >> any problems, though there were some concerns in the past. Please >> evaluate it for inclusion once again. > > Anyone else actually testing this? Any comments?
The speed benefit comes from switching to a stack mbuf except in the case of errors, right? There should still be space for saving 68 bytes though, shouldn't there? Is copying 60 bytes a big difference?
