Re: RAC or Large SMP...?

On 09.10.2008 01:24, mccmx_at_hotmail.com wrote:

> By combining multiple x3850 chassis together we will effectively have
> a NUMA based architecture, which means that there will be some memory
> access which has higher latency. However I see this as much less of a
> penalty that the overhead of RAC inter-instance traffic (both network
> delay and IPC overhead).

I do not have experience with NUMA or RAC - all I can offer is my skepticism. And this tells me that I'd rather not jump to this conclusion too quickly. For example, because Oracle RAC knows the application chances are that the data exchange is optimized for the usage pattern of a RAC. The multi chassis option may sound faster in theory but it may not be in practice. Assuming that it will transfer memory pages of maybe 4k but your Oracle block size is 8k. Now you need two transfers per block where the RAC might have to do just one. This alone can have an impact. Here's another one: Oracle uses large shared memory for SGA - maybe shared mem and NUMA do not play well together.

> This doc explains the architecture pretty well I think:
>
> http://www.redbooks.ibm.com/Redbooks.nsf/RedbookAbstracts/tips0476.html?Open
>
> I haven't made my mind up either way yet but I'm definately favouring
> the multiple x3850 idea over the multi-node RAC cluster.

I'd definitively do some more research or even experimenting.

Kind regards

        robert Received on Thu Oct 09 2008 - 11:43:57 CDT