Maya Mental Ray problem with new Mac Pro

[robinp]

Having bought a new mac pro 2.26 8-core model to render with mental ray in maya, I am extremely disappointed to find it only take advantage of half of the available CPU resources... bummer. Does anyone know much about how hyperthreading works? Is it likely that it is still making use of the 8 physical cores and it just appears that there are 8 idle cores?

Any advice would be welcome because at the moment I'm considering repacking the thing and sending it back...

 

[Pika]

What about Maya for Windows? Did you try to Boot Camp to Windows and test it to see if its the Maya version of OS X problem? I think its a software issue and not an hardware one... because i use softimage to render models with mental ray and it works fine. Of course... i had to use Windows.

 

[Abidubi]

Thats an interesting question. If a program only has 8 threads, and activity monitor shows 16 because it counts the virtual cores... if the program uses all 8 cores to the max what will the 8 virtual cores show? I mean if there are only 8 threads, it can't use more than 8 period. But if the program is fully using those 8 cores wouldn't it show 800% CPU instead of 1600% CPU despite the fact that it is working just as fast as it would had it been using all virtual cores aswell.

My head hurts.

 

[robinp]

I haven't had a chance to install bootcamp yet. I did read a while ago that mental ray is limited to 8 cores on the host computer, which is why this was the first thing I tried... it was a concern. Obviously my new machine shows up as 16 cores to the OS so this could explain why I am having troubles. Maybe should have gone with the quad 2.93 instead.

Will try with bootcamp in due course though. Right now, have to dash out

 

[robinp]

Thats an interesting question. If a program only has 8 threads, and activity monitor shows 16 because it counts the virtual cores... if the program uses all 8 cores to the max what will the 8 virtual cores show? I mean if there are only 8 threads, it can't use more than 8 period. But if the program is fully using those 8 cores wouldn't it show 800% CPU instead of 1600% CPU despite the fact that it is working just as fast as it would had it been using all virtual cores aswell.

My head hurts.

yes this is exactly what I was getting at. Just did a test render on my old quad 2.6 and on my new octad 2.26 (which shows just over 800% CPU use)

quad 2.6 = 1:44 (one min 44 sec)
octad 2.26 = 1:01

so it seems it isn't fully using the power. bugger

 

[robinp]

In fact it seems mental ray performance in maya is scaling linearly with clock speed. None of the speed up with the new architecture that's been seen elsewhere with nehalem and 3D rendering. I think it must be down to not utilizing all '16' cores

 

[Abidubi]

yes this is exactly what I was getting at. Just did a test render on my old quad 2.6 and on my new octad 2.26 (which shows just over 800% CPU use)

quad 2.6 = 1:44 (one min 44 sec)
octad 2.26 = 1:01

so it seems it isn't fully using the power. bugger

I not so sure. It is 43 seconds faster (170% as fast) with only 87% the clock speed. If you had a 2.66 octo, it'd probably be almost 200% as fast as your quad core, which makes perfect sense.

Sounds to me like it's using all the power it has. The other 8 cores show idle because it isn't able to exploit multi-threading, but it wouldn't be faster because then each core would be doing twice as much work for the same number of instructions it can execute (just doing more side by side, but not more work).

 

[scanline]

Seems like a reasonable performance increase to me as well. If you can figure out how to set the processor usage to a single proc, you could see how it is scaling the render times between one and eight procs. There is usually enough overhead that can't be done in parallel that you probably wouldn't get a good result, especially on a one minute render. Try a 4 hour render on the old machine and then see what it does on your new machine.

 

[Abidubi]

There is a processor utility in the developer tools that lets you set the number of processor, and you can disable whichever you like.

http://www.versiontracker.com/dyn/moreinfo/macosx/14358

Another brainstorm... if you force half the processors off, do you force the processors to over-clock too(Turbo boost)?

 

[Dreamail]

Having bought a new mac pro 2.26 8-core model to render with mental ray in maya, I am extremely disappointed to find it only take advantage of half of the available CPU resources... bummer.

I think this is a licensing issue.
IIRC Mental Ray for Maya is only licensed for up to 8 cores. Check with your Maya retailer.

The good news is, if you purchase extra copies of Mental Ray Standalone you should be able to render with all 16 threads.

Another tip the manual suggests in regards to Hyperthreading:
Disable “Auto Render Threads” and set the number of render threads manually.
(or try the opposite - see if it makes a difference).

 

[fluidedge]

sorry am i missing something here?

2 quad core CPUs = 8 cores.

What is this talk of 1600% and 16 cores?

Maya for mac (apart from just plain sucking) is not a 64 bit app and so i doubt there is much in the way of Hypertheading/acceleration. Heck nCloth simulation caching only utilises 1 core.

I'd say a 170% increase in the render speed is pretty damn good. If you move over to a Windows licence of Maya 64 bit you might get a bit more in the way of threading (nCloth cache will still only use 1 core) but i doubt it.

Consider using a renderfarm if you want more power.

 

[Abidubi]

sorry am i missing something here?

2 quad core CPUs = 8 cores.

What is this talk of 1600% and 16 cores?

Maya for mac (apart from just plain sucking) is not a 64 bit app and so i doubt there is much in the way of Hypertheading/acceleration. Heck nCloth simulation caching only utilises 1 core.

I'd say a 170% increase in the render speed is pretty damn good. If you move over to a Windows licence of Maya 64 bit you might get a bit more in the way of threading (nCloth cache will still only use 1 core) but i doubt it.

Consider using a renderfarm if you want more power.

Software sees 4 cores as 8, and 8 as 16 due to hyperthreading. 64bit has nothing what so ever to do with the number of cores a system uses, thats memory.

And activity monitor has always reported CPU usage as a %, with 100% being 1 processor. 16 virtual processors is 1600%

 

[robinp]

thanks for the tips guys, I shall look into these various options and report back. As far as the 170% increase, on the face of it it does seem reasonable, but when considering the 30-40% increase in performance per clock it doesn't.

eg

2.26*8*1.3 = 23.5
2.6*4 = 10.4

23.5/10.4*100 = 226% expected increase in speed

 

[robinp]

There is a processor utility in the developer tools that lets you set the number of processor, and you can disable whichever you like.

http://www.versiontracker.com/dyn/moreinfo/macosx/14358

Another brainstorm... if you force half the processors off, do you force the processors to over-clock too(Turbo boost)?

just tried downloading this, but it redirects to the ADC site and I'm not a member

 

[Zandalus]

just tried downloading this, but it redirects to the ADC site and I'm not a member

You can make free registration

 

[robinp]

unfortunately the CPU preference panel doesn't seem to work quite right with the new CPU's. For a start, it says it is an 'unknown processor' and then when clicking the hyperthreading box, it just switches it to 4 cpu mode rather than actually turning of hyperthreading. Shame.

What is very clear from watching this panel while rendering, is that often 2 threads of the 8 will be running on the same core, which is not a great use of resources

 

[robinp]

ok, to give an idea about what kind of effect this problem is having:

When setting the CPU's to turn off hyperthreading, it actually just switches to 4 cores... there isn't a combination where it will let you have 8 cores and no hyperthreading. In this 4 core scenario, a render took 1:07.

With 8 cores (and hyperthreading having to be on) it took 0:56. During this test there were 3 cores running 2 threads each, and a further 2 cores running only 1. This varies from one time to the next. It would be reasonable to expect that if the 6 threads that were sharing 3 cores were actually distributed across 6 physical cores that the performance would increase by a further 20-30 seconds or so given the 10sec improvement when 2 threads have their own core.

Apple must make a way for us to turn off hyperthreading, or Autodesk should just make sure that mental ray for maya understands the difference between cores and threads

 

[Abidubi]

ok, to give an idea about what kind of effect this problem is having:

When setting the CPU's to turn off hyperthreading, it actually just switches to 4 cores... there isn't a combination where it will let you have 8 cores and no hyperthreading. In this 4 core scenario, a render took 1:07.

With 8 cores (and hyperthreading having to be on) it took 0:56. During this test there were 3 cores running 2 threads each, and a further 2 cores running only 1. This varies from one time to the next. It would be reasonable to expect that if the 6 threads that were sharing 3 cores were actually distributed across 6 physical cores that the performance would increase by a further 20-30 seconds or so given the 10sec improvement when 2 threads have their own core.

Apple must make a way for us to turn off hyperthreading, or Autodesk should just make sure that mental ray for maya understands the difference between cores and threads

With such a slight change in performance, I'm inclined to believe that the processor program is not properly built to represent the new cores properly. When you disable the hyperthreading and it shows 4 cores, it seems much more likely that it is actually running on 8 because you results didn't change much. If it really was 8 threads on 4 CPUs, don't you think you would see a much much lower performance than 8 CPUs? I don't know, just speculating. Try sending a bug report to apple.

 

[robinp]

With such a slight change in performance, I'm inclined to believe that the processor program is not properly built to represent the new cores properly. When you disable the hyperthreading and it shows 4 cores, it seems much more likely that it is actually running on 8 because you results didn't change much. If it really was 8 threads on 4 CPUs, don't you think you would see a much much lower performance than 8 CPUs? I don't know, just speculating. Try sending a bug report to apple.

I'm not sure you understood my previous post. The result with all 8 cores turned on had 6 threads running on 3 cores, so only 2 threads benefitted from all 8 cores being turned on

edit - so basically, OS X (or maya?!) isn't very clever about distributing the workload with hyperthreading turned on

 

[robinp]

just to confirm that the processor pref panel doesn't really know what it's doing with these chips see the below pic. Also, if you un-tick the hyperthreading box on that window, you are left with 8 cores visible on the list (same as with hyperthreading enabled), but the drop down changes to 4 CPU's and if you untick the CPU's in the list, when you turn off 7&8, the drop down list changes to saying 3 CPU's

 

[robinp]

to illustrate what I was saying above, see the image below. What this seems to indicate with the CPU list is that it is actually 4 cores with hyperthreading turned on despite having turned it off.

 

[Abidubi]

No, this really looks to me like the checkboxes are correct, and what is shown in the graphic is that the first 4 CPUs are actually all 8, and the last 4 are the extra "hyperthreading" threads. Where it says CPU 1 Thread A and Thread B, I think that is actually core 1 and 2.

Your own comments support that. You said when you disable hyper, you see CPUs 1,2,3 and 4. When you go to the checkboxes and disable the last 2 (7 & 8) CPU #4 disappears from the graph. That would support that all the hyper threads were already disabled, and you just turned off 2 cores which where being represented by CPU 4 Thread A and B.

It may be either the way it was programed for all processors before hyperthreading, or a error made. Try having hyperthreading on, and disable ONLY CPU 2. In the graph what happens? I would guess CPU 1 Thread B, and CPU 5 Thread B will stop.

Also when you say when running Maya it shows 3 CPUs with 2 threads and 2 with 1. Well if the above is true, then it means CPUs 3 to 8 have 1 thread, and CPU 1 and 2 have a second thread running. That makes absolute perfect sense.

Again, I'm speculating, but I really really really don't think you would get a 200% improvement with such a drop in clock-speed.

Look at this graph taken from "2009 Mac Pro in hand - unboxing etc.". The new 2.26 is not twice as fast as the old 2.66 octo. 16,615 (2.66) vs 18,088 (2.26). And with geekbench, 11397 (2.26) vs 8090 (2.8). You shouldn't be expecting 200%. You are getting 170% and that makes perfect sense considering the new architecture.

 

[jimithing1]

edit - so basically, OS X (or maya?!) isn't very clever about distributing the workload with hyperthreading turned on

This is correct, supposedly it will be corrected in 2010. Maya OSX is supposedly getting a full rewrite.

 

[fluidedge]

yes maya is still a Carbon app. Blame apple for dropping support for Carbon, a Cocoa rewrite is a monumental task. I wouldn't expect it for Maya 2010 as internal betas are already floating around for that.

 

[robinp]

No, this really looks to me like the checkboxes are correct, and what is shown in the graphic is that the first 4 CPUs are actually all 8, and the last 4 are the extra "hyperthreading" threads. Where it says CPU 1 Thread A and Thread B, I think that is actually core 1 and 2.

Your own comments support that. You said when you disable hyper, you see CPUs 1,2,3 and 4. When you go to the checkboxes and disable the last 2 (7 & 8) CPU #4 disappears from the graph. That would support that all the hyper threads were already disabled, and you just turned off 2 cores which where being represented by CPU 4 Thread A and B.

It may be either the way it was programed for all processors before hyperthreading, or a error made. Try having hyperthreading on, and disable ONLY CPU 2. In the graph what happens? I would guess CPU 1 Thread B, and CPU 5 Thread B will stop.

Also when you say when running Maya it shows 3 CPUs with 2 threads and 2 with 1. Well if the above is true, then it means CPUs 3 to 8 have 1 thread, and CPU 1 and 2 have a second thread running. That makes absolute perfect sense.

Again, I'm speculating, but I really really really don't think you would get a 200% improvement with such a drop in clock-speed.

Look at this graph taken from "2009 Mac Pro in hand - unboxing etc.". The new 2.26 is not twice as fast as the old 2.66 octo. 16,615 (2.66) vs 18,088 (2.26). And with geekbench, 11397 (2.26) vs 8090 (2.8). You shouldn't be expecting 200%. You are getting 170% and that makes perfect sense considering the new architecture.

it's a 2.6 QUAD not octad that I'm comparing it to and beside that, I really don't follow your logic. There are clearly some serious flaws with the way the processor preference pane works with these new chips - I guess we can expect a new version with Snow Leopard/WWDC