What's not multi threaded in MR

[pixelvapour]

This has been a long term annoyance in MR.

A distributed render will hang at 99 percent, for say 2hrs out of a 5hr render. If I'm using 5 machines to distribute the render, each is an 8 core MacPro, so 16 hyper threaded cores. The render will be using about 2 cores, on 2 machines, (1 core on each) out of a possible 80 during these 2 hours, so what's not hyper threaded?

This isn't a one off it happens on a regular basis. It's not every scene, but there is something, either materials or combination of materials that's causing the machines to calculate using only a single core.

Basic details are:

Mental ray 3.8
MacPro 2 x 2.93 Quad-Core Intel Xeon
OS X 10.6.4
Qube 6.1

Thanks,
Dave

[RGhost77]

Fast SSS and materials with lightmap shader is threaded by object. So if scene contains one object with this shader only one core is used.

[pixelvapour]

OK, that's handy to know. This scene doesn't contain any SSS or lightmap shaders so I assume there are other factors causing it to drop to one thread. It's only using carpaint and mia materials.

Does anyone know of other single threaded mental ray elements?

Dave

[seb]

i´m not that kind of an expert here but i think thats not a question of multithreading or not.
if you do a dr every core renders one bucket. if it´s finished this core gets the next one and so on. finally there are maybe just two buckets left so only 2 cores can render them. thats just it. had a similar "issue" and the last buckets hang for an additional long time because there was a lot of blurry refraction and reflection stuff in these which you wouldn´t notice while all cores are rendering continuous.

what helped me a lot in those cases is playing around with the task size. dont let mr automaticallly adjust this can speed up rendertimes in those cases a lot!

[pixelvapour]

Yeah I've heard that mentioned before, and it does make sense. It's a pity that mr can't manage buckets more efficiently, all this expensive hardware sitting idle is just wasteful.

I'll look into changing the task size. The problem is that you generally don't discover that your 8000px render has dropped to 1 core until it's been running 20hrs and at 99 percent. Not really the time to cancel it and start tweaking task sizes.

Thanks for your help.

Dave

[steve]

Hi all,

perhaps it's worth to take a look at this thread which is talking about multi-threading limitations in mental ray related to licensing.

I would like to make some comments on the problems you are seeing.

The mental ray core itself provides basically unlimited parallelization capabilities, across multiple CPUs, CPU cores, and even remote 'CPUs' (mental ray Satellites). The actual parallel execution efficiency, like how many tasks or jobs can run simultaneously in a specific rendering, may be reduced though, by things like:

• I/O, like file/stream reading, scene parsing, image file writing, network transport,
• Geometry, pieces that are large and indivisible,
• Textures, if they are large and indivisible (non-tiled),
• Shaders, if they access or share a common resource, like singular pre-computed data or light map,
• Effects, if they have global impact, like deep ray tracing or global illumination, reaching all parts of a scene almost at the same time.

mental ray comes with many kinds of optimizations to improve parallel efficiency, like

• Splitting of large pieces of geometry,
• Tiling of large texture maps,
• Tiling of large frame buffers / output images,
• On-demand loading of geometry only during rendering.

It is important to note, that rendering multi-hosted, also called distributed network rendering, adds more complications to parallel efficiency. Rendering effects like final gathering may not benefit in performance noticeably since they operate in a global context and may share a global resource (that needs to be delivered to and from all remote hosts).

In conclusion, problems with parallelization need to be looked at case by case, to detect where the bottleneck is, and to propose the best ways to work around it.

Please feel encouraged to report any noticeable degradations in performance and parallelization, preferably with details about the hardware (computer systems, processors, memory) and software (mental ray version, OS version).

Best, Steve.