Re: 10g RAC design options

Comments in-line.

ORA600 wrote:
> Hi Daniel,
>
>
> all my comments here come from real life experience and not
> assumptions.
>
> yes, it is true that adding nodes to a RAC cluster is one of the major
> marketing points and the stupendous ability of RAC to accommodate that.
> I myself have configured/added more nodes when applications are written
> well to scale and perform.
>
> Just for the records, we are talking of scalability here and NOT RAC's
> ability to scale. If an application is crappy, come what may, RAC WILL
> NOT scale.

No truer words were ever said. But one should note that the bloody mess likely won't scale on an SMP or MPP machine either.

> i am referring to the fact that adding nodes is not the solution
> *always*. Many a time, adding resources to your existing node(s) have,
> has and will solve scalability issues.

That has not been my experience.

When I see scalability problems with RAC I find the issue is almost always a poorly architected application with every block on the planet being passed across the mem-interconnect.

Though I did find one last year in which the developers were emptying tables with "execute immediate truncate" which stopped it dead in its tracks.

> What i was refering to is the fact that applications (and servers) DO
> NOT scale just because you keep adding nodes. 10g RAC can probably take
> a million node cluster and million Petabyte database but what are going
> to run on it? Solitaire, Doom or one of the new age violence ridden
> games?

Actually the largest 10gR1 RAC cluster ever built, AFAIK, was 128 nodes. And my understanding is that officially Oracle will not support more than something like 64 - 80 (Mark likely knows the actual limit). And you are correct about what one could potentially run on it though I am now looking at someone considering an extremely large cluster for SAP.

Still the issue is what is the difference between 4x4 vs 8x2. Not how many zones one can configure in a fiber switch.

> have you ever tried running a datawarehouse on a 4-node Linux cluster?

Yes.

> if you write your application from scratch on Linux or adopt it for
> Linux, then you have something to argue about scalability and adding
> 1024 nodes to an existing cluster.

Not sure where you're going with 1024 nodes as it doesn't exist.

> have you tried a commerical billing, banking or enterprise
> application(s) on a 4-node Sun/AIX/Tru64/HPUX/Linux cluster and have
> you actually seen what problems these applications face? i am not even
> going to talk about 8-node clusters and beyond.

I am familiar with a rather well known 16 node cluster just over the bridge from me at Amazon.com. Knowing that it, and others work quite well I'm at a loss to understand your point.

Certainly there is a comlexity issue. But properly built they are easily handled. I will be building two 10 node clusters in three weeks: It won't be the largest and it won't cause me any loss of sleep.

> it is acceptable that applications may be badly written and RAC is not
> blame. Amazon's datawarehouse and web applications are written well and
> is RAC adopted.

Exactly. So the issue is NOT RAC ... the issue is badly written apps.

> Having a 64-Node RAC cluster with Web servers, Apps Server and
> concurrent managers on them is NOT a very intelligent thing to do.

A blanket statement to which I can not agree.

I wouldn't be my first choice architecturally but if contracted to build it I have zero doubt about the ability to be successful provided, and you already know the provision, the application was designed well. And by well I mean that if Tom Kyte or Jonathan Lewis looked at the design their blood pressure wouldn't go up more than 1-2 points.

> If you don't like Sun boxes or if it doesn't work according to your
> expectations, you are entitled to your opinion. If you think blindly
> adding nodes is the answer to scalability woes then i think you may
> have some misunderstanding.
>
> regards
> ski

I never think blindly adding anything is the answer. So lets do this with eyes wide open.

Consider this challenge. You build whatever you want using Sun boxes (including storage, support contracts, rail kits, everything) then give me a budget 65% the size of what you just spent and let me build a well designed RAC cluster with zoned fabric switches.

For that 65 cents on the dollar I'll leave your Sun boxes smoking in the dust, will have both DataGuard and Transparent Application Failover, and will retrain your UNIX sys admins in the new operating system to the point that they are competent.

If you decide to take the challenge ... start by looking at the power consumption of the Sun boxes. You'll spend more on the air conditioning, power conditioning, and UPS's than I before we ever even get to the rest of the technology stack because they put out over twice the BTUs as my choice. ;-)

Eyes wide open ....

-- 
Daniel A. Morgan
http://www.psoug.org
damorgan_at_x.washington.edu
(replace x with u to respond)