Here is a response from James Morle on this thread. For those who don't recognise his name, James is the author of the book "Scaling Oracle8i" which is on the recommended list of Advanced Oracle Resources on the Ixora web site. Chapter 1 of that book contains superb explanations of many of factors behind James' and my recommendation of "fewer, faster CPUs". That chapter, Scaling Concepts is one of the sample chapters from his book available on the publisher's web site. Unless you already own a copy of James' book, I've no doubt that reading that chapter, and the other sample chapter, Scalable Transaction Processing will whet your appetite for a visit to Amazon.com!

Not sure about that scenario. It would only apply in the case where the workload was contention free. Any kind of cache coherency operations required between the CPUs will slow the whole process down a good deal more than the relative decrease in context switching gains. With the shared memory architecture that Oracle has, there is a very high chance of some kind of reuse of memory locations between the processes. Parallel Fortran may represent the case well, but Oracle surely does not.

...

I can honestly say, I'm firmly in the "fewer, faster" camp. Running benchmarks on NUMA machines really drives home the impact of coherency operations because it is so exaggerated on those platforms.

The recent Ixora News item Buying Time has provoked some discussion on the question of whether to buy fewer faster CPUs or more slower ones.

It has been pointed out that there is one case in which it might be better to buy more slower CPUs. If an application has a number of highly CPU intensive processes (or threads) greater than or equal to the number of CPUs available, then the CPU usage will be unacceptably high. High CPU usage increases context switching overheads and introduces a risk of latch contention. If for some reason the CPU intensive processes cannot be tamed, then buying a greater number of slower CPUs may give better overall performance, despite the reduced CPU capacity and scalability.