[TriLUG] Hyperthreading and Xeons
kevin at flanagannc.net
Wed Oct 13 18:54:46 EDT 2004
I have an application that is licensed by the processor, and it sees the
HT processor as if it were real, costly....
That said, we did some quick tests, turn HT off, turn it back on, and
for this app it was nearly linear performance gains, YMMV, depending on
applications. This was on Windows 2003 server, which claims to
support/exploit HT fully.
The reading that I did boiled down to, test it yourself, but expect
about 70% + scalability, if you aren't paying for licenses for software
based on "CPUs" then it seems to me that most of the time you'll do well
to leave it enabled. Isn't it frequently better to run 2x as many
jobs/threads even if they take a little bit longer to process? Depends
on the app, your config, etc.
On Wed, 2004-10-13 at 18:01, Jeremy Portzer wrote:
> On Wed, 2004-10-13 at 17:47, John Franklin wrote:
> > One of the big issues with HT is that it looks like SMP to HT-unaware
> > kernels. The difference between SMP and HT is the shared resources on
> > the chip. When moving a process to a second CPU (core), an HT-unaware
> > kernel may blow the CPU cache, move the process to the second core
> > thinking it's a physically separate CPU, and refill the CPU cache. An
> > HT-aware kernel will move the process w/o blowing the cache.
> Would it therefore be prudent to turn off HT if your application is
> memory-intensive? Is it even possible to turn off HT (e.g. BIOS
> setting)? If I can turn it off, I guess I could try to do some
> benchmarks with my particular application...
> I'm using kernel 2.4.21 with Red Hat's modifications, on CentOS 3.3.
If you never see anything that offends you, you aren't living in a free
Kim Campbell - Former Canadian Prime Minister
More information about the TriLUG