Kris Buelens <[EMAIL PROTECTED]> wrote:
>1. The share assigned to a partition is divided by the number of
>processors of the partition. This is the weight the LPAR dispatcher
uses
>when it searches for a "virtual" processor to dispatch on
a real one.
>Example: 2 parttions, each has weight 40, partion 1 has 1 engine, part
2
>has two. So, the LPAR dispatcher has 3 virtual processor to run:
> Part 1 CPU 1 weigth 40
> Part 2 CPU 1 weigth 20
> Part 2 CPU 2 weigth 20
>So, when the system is really busy, the LPAR dispatcher will tend to
pick
>the processor of part 1 more than those of part 2. And thus indeed:
>adding a logical processor to a partition may degrade its performance.
You
>will need to increase the weight of te partition.
>The VM's dispatcher works the same way.
I think that Kris has put his finger on the problem here. I could not see
how taking engines away from minor partitions could adversely affect the
more important ones but this explanation makes it clear.
Instead of having to deal with balancing the weight
needed for overall workload with the weight in the dispatch queues we have
persuaded all concerned to go back to even allocation of processors to
all partitions, (this was our recommendation in the first place but operations
management knew better!).
Thank to all for their input.
Colin Allinson
Amadeus Data Processing