[c-nsp] General switching question regarding load balancing

Keegan Holley keegan.holley at sungard.com
Sat Sep 18 20:25:15 EDT 2010


On Fri, Sep 17, 2010 at 12:17 PM, Darrell Root <darrellroot at mac.com> wrote:

>
> Jay wrote:
> > If a host has two
> > NICs connected to two switches,(the two switches are connected
> > together) and load balancing between them, switch will see the same
> > source MAC from two ports.  How does a switch decide which ports to
> > put in the forwarding table?  Would it switch back and forth every
> > time there is a packet?  Is there any negative effect on the switch
> > when that happens?
>

There are etherchannel modes that take this into account.  If the switchport
is etherchannel there won't be any issues since they will be treated as one
port.  Packets are forwarded and received based on an XOR operation in the
addresses.  If the ports are not etherchanneled they should not have the
same mac address unless manually modified (not a good idea).  There are
etherchannel/trunking modes in most server OS's that will allow you to send
on two ports but to only receive on one to keep from confusing the switch.
 Failover modes are also available.  Also, most server OS's will not allow
you to put the same IP address on an interface that is up and operational.

If you were to have the same IP address on two NIC's with the same mac
address connected to two non-stacked switches, there would be alot of
issues.  Each switch would have a CAM entry for the mac address on it's
respective port that is connected directly to the server.  If a frame were
sent where the destination was on the other switch this would cause the
receiving switch to expire it's entry for the directly connected NIC and
begin sending traffic via the crosslink to the other switch.  This change
would soon be reversed though, as most computers constantly send control
traffic such as mdns and all the smb stuff from M$.  It would technically
work despite the constantly changing CAM entries, if both server nics were
able to receive any return traffic regardless of the source or application.
 It would cause high cpu on some switches depending on how much processing
is done in software.  It's pretty straighforward because the switches do not
know there is a duplicate mac address until traffic is received from the
other switch, hopefully that is minimal.  Overall performance depends on
what's in the rest of the network. For example: everything gets worse if
there are a large number of downstream switches.

Conclusion: Use etherchannel and stackable switches.


More information about the cisco-nsp mailing list