Over the last few days we have been looking at getting a Client Access Array going for our Exchange 2010 setup for basic redundancy and load balancing. I thought I would outline an issue we discovered with using Windows Network Load Balancing in a HP switching environment.
The first is whether to use Unicast or Multicast modes for the NLB Traffic. It is important to remember that Exchange does not care if you use Unicast or Multicast and is entirely dependent on your switching environment. At first we were confused as to which to choose due to the myriad of documentation suggesting either protocol. So we decided to give multicast a go.
After a few minutes our core switch (an 8212zl) started dropping packets for our services VLAN. We shut off the NLB and everything had returned to normal. After a bit of digging we found that ProCurve switches are not compatible with multicast mode NLB due to their inability to have static ARP tables (they can only cache). But this can be remedied somewhat by issuing the following command on some models of switches:
ProCurve(config)# ip arp-mcast-replies
The command is supported on HP ProCurve E8200zl series, E5400zl series, E3500yl series, E6600 series and E6200yl series switches support Microsoft Windows NLB in Multicast mode. Simply ensure that these switches are running K.15.03.0007 or greater firmware.