24 (actual) core Slower than quad

[Jacolath]

Hi Peeps,

I have just set up a new system: 2 x 12core magny cours 32 gig ram. I had many headaches getting it to work probably and eventually ended up chucking win7 and going back to trust xp 64 pro. Overall it is much faster then my old quad (q6600) machines except when it comes to rendering proxies.

After a lot of tearing of hair and many facepalm moments I have isolated it to only proxies. When i render a scene with 3 proxy trees - no lights no gi the 24 renders 21seconds and the quad renders 18seconds. The problem is further compounded when i render with opacity mapped leaves where the 24 core is 3:30 and the quad 1:12. The same test with normal geometry trees the 24 core blows the quad away and renders bout 20% of the time it takes the quad.

I have come to realise that it is probably to do with the 24 cores all trying to pull the proxy from the HD at the same time. I tried to make a ram disk but the results where the same (funny). Probably the same issue with the 24 cores trying to access one ram module simultaneously. But this made me wonder ....

If i set up a 5 disk raid 0 array would I see a dramatic increase in performance? Or would i get the same bottleneck?

Any thoughts?

On a side note the ATI 5870's viewport performance is not noticeably better than the 2 year old 8800gtx cards

[vlado]

Which V-Ray version is this? V-Ray 2.0 is somewhat better in this regard, and it will get better in the next update.

Best regards,
Vlado

[Jacolath]

Wow quick response Vlado.

It is vray sp5 max 2010. I did not want to upgrade to Vray 2 in the middle of a big project. But maybe if there would be a performance increase with this then I may have to do it. Have to deliver high def animation in 2 weeks O.o Without getting too technical how would vray 2 be different?

I also found now with a render of displacement that its only using 20% of cpu power even though I have Dynamic memory limit set to 28000 it only seems to be going up to 10gig. Are there any setting in my hardware that could be affecting this?

Thanks for your time.

[vlado]

No, it's not a hardware problem. Those parts of V-Ray 1.5 are single-threaded and there is some extra thread locking. You could try to turn off hyperthreading (if there is one) and see if it make things better. In any case, it might be worth trying this with the demo of V-Ray 2.0 and see if it works better.

Best regards,
Vlado

[Jacolath]

Proxy:
There is no hyperthreading it is 24 actual cores. I tried vray 2 demo and on the 3 proxy tree test (as above) it was slightly faster 2:40 vs 3:30 but the quad still does it in 1:12 (on sp5). - For interest sake the 24 core stays @ 100% cpu throughout this test.

Displacement:
There is another scene which I tried which is a blow-up of displacement where the quad is also much faster. I tried to render it locally on the 24core and through DR from the quad. Even if I load the 24 core through DR from the quad u can see the core from the Quad go fast and the cores from the 24core go at half pace. The scene assets are on a shared drive which is on the 24core. Also in the IR map pass the cores from the 24c is slower than the quad. I think the Raid hard disk might solve the proxy issue but dont know that it will help with the displacement test. Even if it is an un-threaded process issue why are the quads rendering faster on the same displacement? Treid in Vray 2 also - Please see attached screen grabs of each pc at 8 minutes.

All the other geometry renders much faster on the 24core.

I dont know if there is some problem with the way my cpus access their memory or something but I am at a loss.

[vlado]

Ok; can you tell me what kind of CPUs are they? Also, have you disabled the 3ds Max Bitmap pager? It has very bad performance on many-core machines and affects all scenes with bitmaps. In 3ds Max 2011 you can disable it from the Common tab of the Render Setup dialog, the section "Bitmap performance and memory options", click the "Setup button" and turn off the "Page Large Images to Disk" option. You will need to save and re-load the scene for this setting to take effect.

Best regards,
Vlado

[add101]

What CPUs? AMD Opteron?
Here is a review http://www.anandtech.com/show/2978/a...-6-core-xeon/7
"3dsmax 2010 crashed on almost all our servers." I would have thought server CPUs would be more stable...
Maybe playing with bucket sizes could help?
Or moving to Linux and blender :P (just kidding)

[vlado]

Try disabling the Bitmap pager first...

"3dsmax 2010 crashed on almost all our servers."

There was a bug with the mentalray plugin that crashed 3ds Max on startup on systems with more than 16 logical processors. This was corrected with a hot fix, from what I remember...

Best regards,
Vlado

[Jacolath]

The CPU info u can see in a quick bench I ran - attached image.

The bitmap pager disable seems to have resolved the displacement problem we where having and it now breezes through at about 2 minutes. Thanks a lot Vlado

It did not however work for the proxy issue. When the buckets hits those proxies they come to a virtual stand still although it is still reading 100%. The quad goes quickly through the proxies (total 2minutes for region on its own) even though it is pulling the proxies over the network from the 24core. See attached image.

So I'm still sitting with this proxy issue. I still think it might that the 24 cores is somewhow swamping the single HD but .... that would not explain why the quad cores go twice as fast when they rendering the same scene simultaneously in DR.

Any other ideas?

Maybe I should try and get a raid array and see if it helps but dont know if i will get it in time

Yes I know about the Max MR thing. I had Max 2010 with all the hotfixes then I uninstalled it because I thought it might be the hotfixes causing the problems like with the vrayspawner issue. Currently just removed mentalray.dlz from plugins to allo max to load with 16+ cores. Dont use mental ray :P

[vlado]

It would be nice if we could somehow get the scene for testing and debugging, but for the time being it might be easier to just convert the proxies to meshes, seeing as how you still have free RAM.

Best regards,
Vlado

[Jacolath]

Unfortunately it is not specific to that scene. Any scene with proxies in make the 24 core render slower.

Interestingly in my test scene that only had 3 proxy tree and no lights cameras or GI ... when I make the 3 proxie trees just copies of the same proxy and not instance it render in half the time. Then it is roughly the same speed as rendering with the quad. But i still expect the 24core to be much faster than the quad.

[vlado]

I will look into it, but I'm afraid there isn't much to do about it on your end at the moment. Still, if you have any scenes that we can use for testing, they are most welcome as they will help us to find and fix the issue.

Best regards,
Vlado

[Janosik]

when i changed renderer to MR all my speed problems was disapear when using AMD cpus, MR maxwell fry rendere has no problems only vray has problem with AMD cpus what a surprise ! , it make me think it is not a coincidence ....

[vlado]

<shrug> He didn't say anything about an AMD machine, so not sure how you figured that out. In any case, we are getting a newer AMD machine for testing here, and, like I said before, it would be very useful to get a scene from you that has issues.

Best regards,
Vlado

[Jacolath]

Like i said it is not really scene specific. I can make three proxy trees in a blank scene and illustrate the problem. I will illustrate this more clearly once I have time but for now I have to do everything to try to meet this deadline for next week.

In the CPUBench attachment image earlier in this thread it shows that my system is 24core opteron (the one at the top) .. thats how he knows its AMD I guess.

Thanks for the help so far Vlado.