Re: Oracle 9i RAC vs Hardware clustering (like HACMP)

On 2004-10-26, DA Morgan <damorgan_at_x.washington.edu> wrote:
> JEDIDIAH wrote:
>
>> On 2004-10-20, DA Morgan <damorgan_at_x.washington.edu> wrote:
>>
>>>Mark Clark wrote:
>>>
>>>
>>>>On Mon, 18 Oct 2004 21:14:38 -0700, DA Morgan
>>>><damorgan_at_x.washington.edu> wrote:
>>>>
>>>>
>>>>
>>>>>Praveen wrote:
>>
>> [deletia]
>>
>>>>Worth considering the limitations of RAC too -- range etc.. That said
>>>>h/w clustering and RAC are often complimentary in large scale
>>>>deployments where they each provide a solution to certain
>>>>problems/challenges. Certainly for large projects the aquisition
>>>>costs are rarely a major factor for a blue chip.
>>>>
>>>>
>>>>Mark
>>>>http://www.linxcel.co.uk
>>>
>>>What limitation? In Japan they ran a 10g RAC cluster with 128 nodes.
>>>Do you think you can find an SMP machine that large?
>>
>>
>> Depending on the size of the indivual nodes, SGI probably makes
>> such a machine already.
>>
>> Although, there is a difference between running and running well. The
>> biggest production deployment I've ever heard of was 30 nodes. However, that's
>> been awhile.
>
> I believe it was 128 nodes with 4 CPU Dells making 256 CPUs and that it

        That would be 512 cpus actually.

        That is equal to a single SGI Altix. As cluster tech progresses, so does SMP tech. SGI will be targeting 1024 cpus next. Sun still lags behind a little. However, that cluster would still resolve to a relatively small cluster size with E15Ks (4 nodes) Given what SGI has done, I am sure that Sun is looking to introduce something even bigger (than a 15K).

> scaled at 80% of theoretical. Mark Townsend might wish to weigh in with

        How did that application scale at smaller cluster sizes? Did each node added from #2 on up yield the same boost to performance, or was there a better increase in performance per node for smaller values of the cluster size?

> the details if he can make them public. I noticed an Oracle publication
> not too long ago list 64 nodes which is 1/2 of what I know was proven.
> This is not surprising as I have never had a RAC fail-over as slow as
> what Oracle guarantees either. They seem to be reasonably cautious in
> what they claim: And reasonably so.
>

-- 
        Negligence will never equal intent, no matter how you 
attempt to distort reality to do so. This is what separates         ||| 
the real butchers from average Joes (or Fritzes) caught up in      / | \
events not in their control.