Re: Dell-Oracle-Linux: Anyone else run this...because its not working for us!

Disable spanning tree? Not a bad idea, but how about getting rid of routing altogether?

Now, I don't personally use RAC, and haven't used OPS in many years. But my gut instinct tells me to build the networks thus:

• One or more "public" network interfaces, each on its own switch if possible. These are the networks that your listeners will listen on, and that users (and DBAs) will use to connect to the database (or login to the operating system). Redundant routers might be a nice touch, but may be overkill if you happen to have a high-availability router...
• At least two interfaces (in each node, of course) for the "private" or "interconnect" networks. Each network should (must) have its own switch, and ideally should NOT be connected to a router. Unless your RAC nodes are (many) miles apart , routing should not be necessary for these networks. At best routers do nothing. At worst, they add latency, provide opportunities for DOS attacks, and give you a number of neat ways to shoot yourself in the foot. Or both feet. And, of course, failure to use switches on these networks is *known* to cause trouble...

My goal would be to keep things as simple and "foolproof" as possible -- in my (perhaps misguided) view, this includes ensuring that only RAC traffic ever traverses (or *can* traverse) the interconnect network(s). I suppose you might have to bend these rules to monitor performance on the switches, unless your switches allow this to be done out-of-band.

Perhaps the RAC gurus out there will be kind enough to tell me whether I'm completely off-base here... (Actually, I have little doubt they will. ;-) )

Cheers,
-- Mark.

時期精霊 wrote:

> Pete Sharman wrote:
>
>> Yep, as you found out the hard way crossover cables aren't
>> supported. Too many problems that magically disappear when you
>> switch to a switch (if you get what I mean!)
>>
>>
>
> We strongly recommend the desactivation of spanning-tree for cluster
> network. (as the architecture should not need spt)
>
> on cisco:
> int fa0/1
> spanning-tree fast
>
> will do it.
>
> Just a little tip, but we had several problems before we found out.
>
> Regards
>
> Kuon

--
http://www.freelists.org/webpage/oracle-l