On Friday 22 September 2006 09:36, Ian mu wrote:
> First thing I tend to do is run something like prime95 or something
> else that will max a virtual cpu and see what it shows.
> [...]
> Typically you won't get much past
> an extra 10/15%, probably be even less with gameservers.

I benchmarked mprime on HT and got something like
running 1 mprime instance used 150ms per iteration (like without HT),
running 2 mprimes instances at the same used each 270ms per iteration.

so,
calculating 2 iterations using one mprime instance (anyway weather HT on
or off) takes 300ms
calculating 2 iterations using two mprime instances on HT takes 270ms
about 10% more throughput.

> Main thing is not to think of each virtual processor as a full cpu,
> even though either one can run at 100% in a sense, but not both at
> the same time, other things are happening. It's worth doing a bit of
> reading on hyperthreading on the net which can explain the details a
> lot better.

Well because It was interesting for me I did above benchmarks again
using different nice levels for both mprime and BTW I discoverd another
Problem which lets me think more bad about HT now:

If you run 2 processes at the same time with different nice levels then
the behavior is not as you would expected on a single_CPU or a
real_multi_core!

Here I posted that more exactly:
http://mersenneforum.org/showthread.php?p=87769#post87769

Dunno weather its a problem of my system (maybe someone could repeat
such benchmark).
For me this means that processes which are usually niced started by cron
etc. got more bad behavior on HT as without HT and I dont see how you
could get around this.

So I think if you still want to use the little speed advantage of HT on
a gameserver then you should watch more carefully your processlist and
cron config even you didnt cared much about niced_19 processes so far.

cu,
Rudi

_______________________________________________
To unsubscribe, edit your list preferences, or view the list archives, please 
visit:
http://list.valvesoftware.com/mailman/listinfo/hlds_linux

Reply via email to