[nsp] OSPF load balancing

[Robert A. Hayden]

You could enable per-flow load balancing, but with two 1.5mb links, they
may not be large enough to achieve proper balancing.

On Thu, 14 Nov 2002, Stephen J. Wilcox wrote:

> cool. the only thing to watch now is that individual flows will split over each
> route, this might be okay however you might see degraded performance if packets
> arrive out of order - a particular problem if you have realtime traffic. you'll
> get this happening if the two circuits have differing latency or when they start
> getting full and queuing data such that different amount of jitter is
> introduced.
>
> Steve
>
> On Thu, 14 Nov 2002, Ryan Roylance wrote:
>
> > Thanks to another suggestion from the list I turned on ip cef, then on
> > each frame interface did ip load-sharing per-packet.  Everything worked
> > perfectly, CPU stayed down, the traffic is flowing along nicely.  Thanks
> > for the help.
> >
> > ryan
> >
> > -----Original Message-----
> > From: cisco-nsp-admin@puck.nether.net
> > [mailto:cisco-nsp-admin@puck.nether.net] On Behalf Of Stephen J. Wilcox
> > Sent: Thursday, November 14, 2002 13:18
> > To: Ryan Roylance
> > Cc: cisco-nsp@puck.nether.net
> > Subject: Re: [nsp] OSPF load balancing
> >
> >
> >
> > On Thu, 14 Nov 2002, Ryan Roylance wrote:
> >
> > > I have a design with 2 routers and two frame relay connections (1.5)
> > >
> > >
> > >          172.17.0.0/16       RA-----------a----------------RB
> > > 172.18.0.0/16
> > >                              RA-----------b----------------RB
> > >
> > > I am running a single OSPF area utilizing the multiple path equal cost
> >
> > > load balancing.  RA and RB are both Cisco 2621 routers.
> > >
> > > The question is this, by default the equal cost load balancing is per
> > > destination, so Im not getting a perfect balance,  I have a couple of
> > > servers on each side that see heavy traffic and they keep on ending up
> >
> > > on the same link which defeats the purpose of having two lines.  When
> > > I switch to per packet load balancing the CPU on the routers goes way
> > > up (from 8% over 5 minutes to 25% over 5 minutes) and the performance
> > > drops ( I went from capping 1 at 1.5 to pushing 400 on each)  What
> > > solutions have people used in the past to make things like this
> > happen?  Are the
> > > 2621's underpowered for what Im trying to do?
> >
> > thats the trouble with per packet .. what you actually do is turn off
> > the hardware switching and force it all onto the main CPU
> >
> > the 2621s are fine, the trouble is the limitation of the routing
> > protocols
> >
> > my suggestion to you would be to manually influence the traffic, you can
> > either do this with a few static routes for the high volume servers or
> > if you want to complicate your config you could advertise more specifics
> > for your servers and then use an offset list to influence the ospf route
> > to each specific
> >
> > the simplest way is the static routes, in the event of a link failure
> > you'll still fall over to the alternate route but you will achieve the
> > traffic balancing you need..
> >
> > Steve
> >
> > >
> > > thanks
> > > ryan
> > >
> > > Ryan Roylance
> > > Technical Operations Manager
> > > OpSource
> > >
> > > _______________________________________________
> > > cisco-nsp mailing list  real_name)s@puck.nether.net
> > > http://puck.nether.net/mailman/listinfo/cisco-nsp
> > > archive at http://puck.nether.net/pipermail/cisco-nsp/
> > >
> >
> > _______________________________________________
> > cisco-nsp mailing list  real_name)s@puck.nether.net
> > http://puck.nether.net/mailman/listinfo/cisco-nsp
> > archive at http://puck.nether.net/pipermail/cisco-nsp/
> >
> > _______________________________________________
> > cisco-nsp mailing list  real_name)s@puck.nether.net
> > http://puck.nether.net/mailman/listinfo/cisco-nsp
> > archive at http://puck.nether.net/pipermail/cisco-nsp/
> >
>
> _______________________________________________
> cisco-nsp mailing list  real_name)s@puck.nether.net
> http://puck.nether.net/mailman/listinfo/cisco-nsp
> archive at http://puck.nether.net/pipermail/cisco-nsp/
>
>