[lug] Re: q: multithreading on smp weirdness

[karheng at softhome.net]

> 
> I forgot to mention, on this machine without SMP APIC enabled: It is 
> probably a fair test if it is not under a large hardware IRQ load, e.g., 
> not huge disk access, any IDE drives have DMA enabled, and network is not 
> under huge load. If any IDE drive does not have DMA enabled, or if 
> networking is saturated, or if any other device could starve the 100 Hz 
> hardware IRQ rate of the 2.4 kernels, then it is not a fair test. Even 
> when fair it is possible that threads will suffer since hardware will 
> always force any required driver handlers to constantly fill cpu#0 cache 
> with it's own data (assuming there is enough pressure on cache that 
> software and hardware drivers compete). Most of the time a machine like 
> this which is not a heavily loaded machine is at least ok with this. One 
> of the things about 2.6 kernels that is supposed to be more efficient is 
> the ability to use a higher IRQ rate than 100 Hz, but I have not looked 
> into exactly what was done, or whether it is purely software IRQ or 
> hardware IRQ also. 
> 

interesting fact... didn't realize they were
changing IRQ freqs even... for all i know,
the timer chip might already be oscilating at
its max.. :) 

that server is actually a borrowed unit from
the marketing dept of another organization for
our evaluation purposes. 

so it doesn't really have much load other than
our tests & evaluations.
otherwise, i almost always make sure no one has
any jobs or scheduled jobs on the servers when
i do my tests. 


rgds, 

kh