[KLUG Members] SMP and ACPI in 2.6.12

[Adam Tauno Williams]

> > The bigger question is, why would you want to enable hyperthreading?
> > Have you properly tested your application performance? With the
> > applications we run our software on it is anywhere from 7% to 18% slower
> > with it enabled than disabled (on 3.4Ghz XEON CPU's). Its particularly
> > brutal on database servers.
> The servers in question are running Sendmail and from what I've read 
> Sendmail benefits from hyperthreading.  No, I haven't benchmarked the 
> difference, but I'm planning on doing so now that you mention it because 
> it sounds like fun. :)  I'll make sure to post the results to the list.
> Doing a quick search online about Linux and hyperthreading got me to this 
> page:
> http://www.2cpu.com/articles/41_4.html
> It seems to show that (at least in this benchmark) that MySQL benefits 
> quite a bit from hyperthreading.  What databases are you running that you 
> see a slower performance?
> Also, are there any guides or benchmarks that you're aware of that make 
> suggestions as to using hyperthreading with different applications?

>From what I've read about HT, I believe -
1.) You benefit most from HT if you have an HT aware schedular.  This
should be true for recent 2.6.x kernels but probably is not true of any
2.4.x kernel,  so a 2.4.x box may actually be slower.  The HT aware
schedular avoids creating cache-misses by recognizing that CPU0 & CPU1
and CPU2 & CPU3 share the shame cache (for example).  [ this is called
'sibiling' CPUs if I recall correctly ].
2.)  For non-threaded single-process services HT may be slower;  but how
many of these still exist?  Apache forks and most RDBMS and DSA servers
use a thread-pool of workers.
3.) Some benchmarks I've seen claim that HT creates a drop in I/O
throughput.  Personally I doubt this;  more actually-concurrent work
creates a heavier load on the elevator so an increase in I/O latency is
a natural event, you need to tune or buy-more-RAM if this is a problem.

I've never tested a single-processor HT box.   But some primitive tests
on my HT servers (all dual-CPU) showed about a ~5-8% performance
increase.  These tests were not controlled enough to be "scientific" but
ruled out any dominant slow-down effect.