Re: Workaround for wrapping loadaverage

[Con Kolivas]

Herbert Poetzl wrote:
> On Wed, Nov 10, 2004 at 08:49:41AM +1100, Con Kolivas wrote:
>
> > Herbert Poetzl wrote:
> >
> > > but I agree that a higher resolution would be a good
> > > idea ... also doing the calculation when the number
> > > of running/uninterruptible processes has changed would
> > > be a good idea ...
> >
> > This could get very expensive. A modern cpu can do about 700,000 context 
> > switches per second of a real task with the current linux kernel so I'd 
> > suggest not doing this.
>
>
> hmm, right it can, do you have any stats about the
> 'typical' workload behaviour? 

How long is a piece of string? It depends entirely on your workload. On 
a desktop machine just switching applications pushes it to 10,000. 
Basically you end up making it an O(n) calculation by increasing the 
overhead of it (albeit small) proportionately to the context switch load 
which is usually significantly higher than the system load.

> do you know the average time between changes of 
> nr_running and nr_uninterruptible?

Same answer. Depends entirely on the workload and to whether your 
running tasks sleep at all or not (hint - most do). While it will be a 
lower number than the number of context switches, it potentially can be 
as high with just the right sort of threads (think server, network type 
stuff).

> TIA,
> Herbert

Cheers,
Con

Attachment: signature.asc
Description: OpenPGP digital signature