Re: [announce] [patch] ultra-scalable O(1) SMP and UP scheduler

[Davide Libenzi <davidel _at_ xmailserver.org>]

On Sat, 5 Jan 2002, Alan Cox wrote:

> > > In fact it's the cr3 switch (movl %0, %%cr3) that accounts for about 30%
> > > of the context switch cost. On x86. On other architectures it's often
> > > much, much cheaper.
> >
> > TLB flushes are expensive everywhere, and you know exactly this and if you
>
> Not every processor is dumb enough to have TLB flush on a context switch.
> If you have tags on your tlb/caches it's not a problem.

Yep true, 20% of CPUs are dumb but this 20% of CPUs run 95% of the Linux world.



> > Again, the history of our UP scheduler thought us that noone has been able
> > to makes it suffer with realistic/high not-stupid-benchamrks loads.
>
> Apache under load, DB2, Postgresql, Lotus domino all show bad behaviour.
> (Whether apache, db2, and postgresql want fixing differently is a seperate
>  argument)

Alan, near to my box i've a dual cpu appliance that is running an mta that
uses pthread and that is routing ( under test ) about 600000 messages / hour
Its rq load is about 9-10 and the cs is about 8000-9000.
You know what is the scheduler weight, barely 8-10% with the the 2.4.17
scheduler. It drops to nothing using BMQS w/out any O(1) mirage.




- Davide



-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo _at_ vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/