[Sidefx-houdini-list] mantra in H9 - multiprocessor

Jed Schwartz jed at jedschwartz.com
Mon Oct 8 12:32:52 EDT 2007


We had problems earlier in the summer when switching production over  
from H8-9 with mantra getting killed upon initiation of a render-   
With some help from Mark E. we determined that the crux of the  
problem was the RHEL distribution we were running-

I was able to get multi-threading working on our newer systems by  
disabling MMX-:

     % setenv HOUDINI_DISABLE_MMX 1	# csh
     % export HOUDINI_DISABLE_MMX=1	# bash
---------------------
 From Mark:

"I know that for an older release of RH Enterprise, we had some serious
problems with threading that didn't occur on other distributions.  The
problems were primarily with using the SSE/MMX instructions, so by
disabling use of these instructions, you might be able to get
multi-threaded rendering working. I had assumed that these issues had  
been addressed in more recent versions of RHEL.  But I'm not sure.."
--------------------------

You can check out the entire thread on the H9 beta forum "mantra  
problems"


However this did NOT work on our older systems running the RH  
distribution-

We have since switched over  to ubuntu and the render problems on ALL  
systems are gone- In fact we far prefer ubuntu to RH in every way- If  
you are running RHEL and having problems with H9, I would strongly  
recommend giving ubuntu a try-

Hope this is helpful!

Jed-


On Oct 8, 2007, at 11:35 AM, Dan Schneider wrote:

> We have been encountering similar issues in H9. I don't think it  
> has to do with HT either. It occurs with the default max proc  
> checked on one of our quadcore xeon machines, but not on the duo  
> core xeons.  Issue ranges to mantra crashing all together. But, if  
> 2 or less threads are used it goes fine, as it has been mentioned.  
> Only when using over 2. What build is everyone using? We are using  
> 9.0.719.
>
> Larry Giunta <larry at gcreativestudios.com> wrote: actually, I'm  
> simply clicking off max proc and
> manually entering thread numbers ........
> so I'm not manually changing any options.
>
> No hyperthreading enabled either -
>
> However, here's something curious, Mark E.
>
> If I copy and paste an output node from H8 into
> H9, then specify 4 processors, the render tile
> just blast up as quickly as expected and
> the render completes much faster than the
> H9 output driver in the same houdini session.
>
> So, it seems to point to some setting or option
> within the H9 output node.
>
> Here are the times I get now. Both rendering
> within H9 ( same object, etc) , but with the first using a pasted H8
> output node.
>
> ------------------------------------------------------
> 4 thread H8 output node pasted into H9
> 1.34u 0.80s 5.76r
>
>
> 4 thread H9 output node in H9
> 1:08.70u 0.55s 17.80r
> ------------------------------------------------------
>
>
>
> On Oct 5, 2007, at 4:46 PM, Mark Elendt wrote:
>
>> On Friday Oct 05 at 16:12, Larry Giunta wrote:
>>> So, we've seemed to narrow down our slow mantra render problem
>>> in H9 to multiprocessor rendering.
>>
>> You are using the new -j option?  Not the -n?
>>
>> By default mantra will be using all your processors.  If you  
>> happen to
>> be using the old -n multi-host way of adding machines, you'll in fact
>> be running 16 threads on a 4 processor system (using -n4).
>>
>> Just a sanity check...
>>
>>> If we render on only one proc, render times seems as fast as
>>> expected.
>>> However, with each increase to the thread count our render times get
>>> increasingly slow.
>>> The system monitor does show the appropriate number of processors
>>> working.
>>>
>>> On the same machine multiprocessor rendering works just fine in
>>> H8, but
>>> in H9 we get the following time increase with each thread added.
>>>
>>> We're running 9.0.739 on FC4
>>>
>>> Wondering if anyone else is getting similar results  ......... or if
>>> this is just unique for us.
>>>
>>> Larry
>>>
>>> -----------------------------------------------
>>> 1 thread
>>> 9.01u 0.38s 9.02r
>>>
>>> 2 thread
>>> 15.15u 0.42s 7.77r
>>>
>>> 3 thread
>>> 56.13u 0.56s 19.02r
>>>
>>> 4 thread
>>> 2:13.80u 0.64s 33.96r
>>> -----------------------------------------------
>>
>> _______________________________________________
>> Sidefx-houdini-list mailing list
>> Sidefx-houdini-list at sidefx.com
>> https://lists.sidefx.com:443/mailman/listinfo/sidefx-houdini-list
>
> _______________________________________________
> Sidefx-houdini-list mailing list
> Sidefx-houdini-list at sidefx.com
> https://lists.sidefx.com:443/mailman/listinfo/sidefx-houdini-list
>
>
>
> ---------------------------------
> Fussy? Opinionated? Impossible to please? Perfect.  Join Yahoo!'s  
> user panel and lay it on us.
> _______________________________________________
> Sidefx-houdini-list mailing list
> Sidefx-houdini-list at sidefx.com
> https://lists.sidefx.com:443/mailman/listinfo/sidefx-houdini-list




More information about the Sidefx-houdini-list mailing list