On Mon, Mar 01, 2004 at 03:18:13PM +1100, David Burren wrote:
> 
> "Toru Nishimura" <locore32@gaea.ocn.ne.jp> wrote:
> >
> > His lab tested HT servers and concluded _the_ HT SMP Linux kernel
> > implementation scales actually worse, at least no better, than
> > plain SMP kernel.  It sounded HT is rather waste of transistor
> > and electricity.
> 
> My understanding is that HT can be beneficial if you have a model
> where you can avoid having threads from separate processes running
> on the same physical CPU.
> 
> Given that this requires the use of multithreaded applications in
> your performance critical path to get any advantage, and a scheduler
> that distinguishes between virtual and physical CPUs, then without
> both of those yes I'd agree it's a waste of time.
> But I'm sure there's an application for it _somewhere_... ;)
> 
> Does anyone know of a scheduler (in any OS) that is HT-aware in
> this way?

Even without all of this, I can get 10% performance speedups in compiling
/usr/src (-j) on an HT enabled 2 cpu box (so 4 total cpu of 2 real, 2 virt) vs
not enabling HT. At that point 10% is 10%. Not great, but I'm not gonna
throw it away either..

James