On 2015/09/18 20:18, David Gwynne wrote:
> 
> > On 18 Sep 2015, at 6:17 pm, Stuart Henderson <[email protected]> wrote:
> > 
> > On 2015-09-18, David Gwynne <[email protected]> wrote:
> >> this lets pf embed the state id into the mbuf as a "flow id" so
> >> other subsystems can use it. eg, trunk can pull it out and use it.
> > 
> > I like this but it does change the path distribution. Previously all
> > flows from host A to host B were bound to a single path, now they are
> > spread across paths. This is good for making fuller use of paths but can
> > make fault diagnosis harder.
> > 
> > I tried to work out the best way to make this optional when I sent my
> > earlier L4-hash diff (using PF states for this has similar results
> > and is far more elegant) but didn't settle on an approach. Switches
> > doing this usually have a single global setting (e.g. sysctl), which
> > seems a bit of a blunt instrument but would be easier to apply to areas
> > other than trunk (e.g. multipath routing). We could use an ioctl if
> > ifconfig(8) isn't full already, though I don't think we actually need
> > any more options than "L3-bound" and "per-flow" so using the existing
> > link0 scaffolding would be an easier way to do this per-trunk but
> > I didn't really get a feel for whether people thought that was
> > good enough.
> 
> the only "fault" ive experienced with hashing algorithms is that they all 
> tend to make the wrong decision, so fixing it has been rotating through their 
> different options till one sucks less than the others.
> 
> is that what you're arguing for here?
> 
> dlg

I meant faults like bad fibre, media converter, switchport, metro
ethernet circuit hanging off an interface, etc. In those cases it's
useful to be able to make it go over a different path by switching
source IP to track things down.

Also sometimes you can get insight into a problem by noticing
that connections from host X in a subnet to host A have problems,
but from Y to A don't, though this more applies where the person
noticing the problem isn't the person who is running the system
doing the load balancing.

Reply via email to