[prev in list] [next in list] [prev in thread] [next in thread] 

List:       linux-smp
Subject:    Re: Problems with an Tyan Thunder i860
From:       Bogdan Costescu <bogdan.costescu () iwr ! uni-heidelberg ! de>
Date:       2002-11-25 15:55:53
[Download RAW message or body]


To FAQ maintainers: Could something related to HT be added to the FAQ ?

On Mon, 25 Nov 2002, Michael Mayer wrote:

> The machine is doing quite well, but I've got problems with
> Hyperthreading.

This was already covered on this list, the beowulf list and probably 
others. It mainly comes from the misunderstanding of what HT provides: it 
does _not_ add a new full processor, but just adds a way of better using 
the processor's internal pipes. So for example two processes, out of 
which one is doing only integer arithmetics and the other only floating 
point arithmetics, could use simultaneously the two different (integer and 
FP) pipes of the CPU; they would appear as running on different processors 
because there is no overlap in what they need. However, as soon as they 
start competing for the same resources (like the FP process would start 
doing integer too) they will have to be executed serially.

> If I turn on Hyperthreading two parallel jobs last around 47 secs, too.

... but here you were just lucky, the scheduler put the 2 jobs on two 
different _physical_ CPUs. But it could happen worse, as you describe...

> I also see some problem of Process routing through the four CPUs when
> starting only two processes. Sometimes they stick to one and the same
> physical CPU and therefore lasting 3 times longer too... 

... where the two jobs were scheduled on the same physical CPU. This is 
exactly the explanation for your problem...

> But the crucial point is when executing four of this jobs in parallel,
> the CPU time should double or if Hyperthreading has any positive effect
> on it, it should take any factor smaller than two longer. The result of
> executing four jobs is lasting 3 times longer.

... which again happens because of poor scheduling. Task switching takes 
some time, flushing pipes and caches also.

> PS: The machine is runing for now with hyperthreading disabled ;(

Well, keep it this way :-)
Especially now that Intel has included HT in its desktop CPUs, you should 
be able to find reviews of what HT could bring (or take); although all 
that I've seen are testing with Windows, the principles should apply to 
Linux as well.

-- 
Bogdan Costescu

IWR - Interdisziplinaeres Zentrum fuer Wissenschaftliches Rechnen
Universitaet Heidelberg, INF 368, D-69120 Heidelberg, GERMANY
Telephone: +49 6221 54 8869, Telefax: +49 6221 54 8868
E-mail: Bogdan.Costescu@IWR.Uni-Heidelberg.De



-
To unsubscribe from this list: send the line "unsubscribe linux-smp" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
[prev in list] [next in list] [prev in thread] [next in thread]