RE: gratuitous arp and bad mac

Paulaner

Are you implementing any Layer 2 Switch Fault Tolerance? Are your Teamed
NICS plugged into two different switches that logically makeup a single VLAN?
Also are your NODES plugged into two different switches that logically
makeup a single VLAN? Microsoft recommends all NODES in a cluster be
attached to the same physical switch and adapter teaming is support on the
public network only but also NOT recommened in a cluster.

Regards

CT

"Paulaner" wrote:

>
> I'm troubleshooting a problem on an Active/Active Win2k SQL 2000
> cluster. We have 2 instances of SQL installed and each instance is on
> a separate node. Lets call them node1, node2, sql-a, and sql-b. In
> this example sql-a is on node1 and sql-b is on node2. The servers are
> connected to a Cisco switch.
>
> A client PC is accessing data regularly through an odbc connection to
> sql-a. This works fine for a while, but after 20 mins or so it stops.
> I can no longer ping sql-a. I can ping node1 and node2 just fine.
>
> I looked at the arp table (arp -a) and found that the mac address for
> sql-a was now matching the mac for node2. So I flush the arp cache
> (arp -d *) and ping sql-a again, it works! The sql-a mac is correctly
> set to node1 again. The odbc app is back up and running.
>
> About 15 minutes later, the problem happens again. Same symptoms -
> wrong mac in the arp table. I checked and the sql-a and sql-b
> instances are running on the correct nodes, there has not been any
> failovers.
>
> So from what I can tell, someone is updating the client PC arp table
> with the wrong mac address. I suspect that the failover from a few
> days ago may be contributing.
>
> I'm guessing that this is happening as a bogus gratuitous arp message.
> I understand the cluster service will send these out upon failover,
> but it looks like it is happening now without a failover event.
>
> The other reason I believe this is a bad GARP is because we added the
> mac as a static arp entry, and even the static value was overwritten
> with the bad mac.
>
> I also wonder if the MAC address logic is confused because the server
> is using teamed nics. The teaming driver (dell) requires a mac
> address to be created and put into the config settings. I wonder if
> the cluster is reading the wrong info here.
>
> I have checked that we have unique mac addresses, and all the other
> settings, so I suspect I have found some combination of events to
> trigger an unexpected 'windows cluster feature'.
>
>
> Anyway - I'll take any suggestions from IP & cluster gurus out
> there... Thanks.
>
>
>
>
>
>
.

Relevant Pages

  • Re: Explain why change from unicast to multicast prevents port flooding
    ... > - Just consider the cluster NIC for now ... address and I think the MAC address as well. ... > 4 Router sends ARP request ... > cluster host to the switch have their MAC addresses masked ...
    (microsoft.public.windows.server.clustering)
  • gratuitous arp and bad mac
    ... Lets call them node1, node2, sql-a, and sql-b. ... I looked at the arp table and found that the mac address for ... sql-a was now matching the mac for node2. ... I understand the cluster service will send these out upon failover, ...
    (microsoft.public.windows.server.clustering)
  • Re: Cluster failover - network level
    ... to the MAC address anymore, and clients already get timeouts/discards, so ... the packet which then the switch shouuld see that the MAC address was sent ... But then the old active isn't listening on that port ... > cluster after failover. ...
    (microsoft.public.windows.server.clustering)
  • RE: WLBS and VLAN
    ... Dedicated IP does not have to be entered on the cluster NIC. ... When MaskSourceMAC is enabled, ... switch continues to see a unique source MAC address. ... switch's requirement that each port be associated with a unique MAC address. ...
    (microsoft.public.windows.server.clustering)
  • mac address issue
    ... I'm seeing two different mac addresses ... it's arp cache, but the switch plugged into that port doesn't. ...
    (Security-Basics)