Kernel Benchmarks With P4+SMP+SMT?

Kernel Benchmarks With P4+SMP+SMT?

[Justin Piszcz]

Has anyone performed any benchmarks with:

No SMP w/HT?
SMP w/HT?
SMP + SMT w/HT?

[ ] Symmetric multi-processing support
[ ]   SMT (Hyperthreading) scheduler support

   x SMT scheduler support improves the CPU scheduler's decision making
   x when dealing with Intel Pentium 4 chips with HyperThreading at a
   x cost of slightly increased overhead in some places. If unsure say
   x N here.

I'm tempted to try SMT and benchmark these sometime but I am asking the 
list if anyone has already done this first.

Question: "slightly increased overhead in some places."

What type of workloads would exhibit such overhead?

Would this option (SMT) be recommended for a desktop or server machine?

Are there any white papers or documentation I can read about this option?

Thanks.

Re: Kernel Benchmarks With P4+SMP+SMT?

[Bill Davidsen]

Justin Piszcz wrote:
> Has anyone performed any benchmarks with:
> 
> No SMP w/HT?
> SMP w/HT?
> SMP + SMT w/HT?
> 
> [ ] Symmetric multi-processing support
> [ ]   SMT (Hyperthreading) scheduler support
> 
>   x SMT scheduler support improves the CPU scheduler's decision making
>   x when dealing with Intel Pentium 4 chips with HyperThreading at a
>   x cost of slightly increased overhead in some places. If unsure say
>   x N here.
> 
> I'm tempted to try SMT and benchmark these sometime but I am asking the 
> list if anyone has already done this first.
> 
> Question: "slightly increased overhead in some places."
> 
> What type of workloads would exhibit such overhead?
> 
> Would this option (SMT) be recommended for a desktop or server machine?
> 
> Are there any white papers or documentation I can read about this option?

I run SMT on all my HT uni systems. Depending on what you do it can help 
up to 30% (kernel build) or just enough to measure. This is one of those 
"it depends" things, I bet there are loads which run better without, and 
there is a tad of overhead in the SMP kernel locking.

If you run SMP, you have that overhead anyway, so I doubt it hurts.

YMMV

-- 
bill davidsen <davidsen@tmr.com>
   CTO TMR Associates, Inc
   Doing interesting things with small computers since 1979