[AG-TECH] inquiry on the hardware specs...

[Frank Sweetser]

On Wed, Oct 22, 2003 at 02:28:40PM -0500, Justin Binns wrote:
> Our past experience with HyperThreading indicates that it does help under
> Windows, but that the Linux scheduler is not hyper-threading friendly, and
> it can actually *decrease* performance (that is, the scheduler in Linux
> already does a pretty good job of keeping the processor(s) busy, having the
> processors themselves double-scheduled doesn't help any, it just creates
> contention for various critical memory regions).  We haven't really played
> with it exhaustively for a while, though - more recent kernels (particularly
> the 2.6 series) may be able to take better advantage of HT technology.

Yes, future versions will handle HT much better.

The problem is that the 2.4 kernel sees each of the four HT cores as a seperate
processor, when in reality it's still two CPUs.  This means that sometimes, the
linux scheduler will schedule two processes on two different HT cores on the
same CPU and and up forcing one of them to block on the other one, even if
there is a physically seperate CPU with both cores sitting idle.  The 2.6
kernel (I've read in the Changelog) is aware of the distinction, and so will
try to spread load around physical processors before it starts doubling up the
HT cores.

-- 
Frank Sweetser fs at wpi.edu
WPI Network Engineer