We have recently setup deadline on our renderblades to render Maya, Max, AE and Krakatoa. Each render blade is minimum dual hex 2.66ghz with solid state drives, 48gb ram and running Win 7 pro 64bot. They currently have multi threading enabled presenting 24 cores and all windows aero turned off.
My question is to do with concurrent tasks. We seem to be able to bump concurrent tasks on some jobs beyond 5 but occasionally we have a blade hang for no apparent reason. Would we be better off turning off the multi threading so only the 12 actual cores are presented to deadline?
Any tips for configuring render machines for these apps would be greatly appreciated. BTW the minimum spec is as above and the maximum spec takes the hex cores to 3.46ghz with same ram and SSD.
Without trying to cover the complete range of applications, here are some thoughts.
When rendering Krakatoa, using HyperThreading helps, but rendering on physical cores is faster. 24 cores will render marginally faster, but if you have enough memory to fit two or more copies of Krakatoa, I would try to run concurrent tasks - you will achieve significantly more with that. Note that 24 GB RAM can fit nearly a billion particles with defaults settings (Position, Color, Lighting, Density channels), and probably half a billion with more sophisticated setups. 8 cores seems to be the sweet spot for Krakatoa 2, above that the performance curve flattens quite fast - it does get faster, but not by as much as you would like.
It is quite possible that the same applies to most renderers in 3ds Max and Maya (mental ray, VRay). 3ds Max Scanline won’t even render on 24 cores - it will crash if you have more than 16. So if memory allows, running at least two concurrent tasks is probably a good idea.
I personally have a 2 years old i7 workstation with 4 cores and HT on and it renders in Krakatoa as fast as a real Dual Quad Xeon from 3 years ago. So I would actually keep HyperThreading on and try to squeeze in more concurrent tasks.
I would not push Concurrent Tasks farther than 4 myself, but I don’t know the size of your jobs, so that’s a bit subjective.
In Deadline 5.1, you will also be able to render multiple jobs on the same machine by launching multiple slaves, but that will require additional Deadline licenses. The benefit is that you can have high-CPU/low bandwidth jobs (typically 3D and simulation) processing parallel to low-CPU high-bandwidth jobs (typically 2D comps) on the same hardware, potentially leveraging all the power and bandwidth better than multiple tasks of the same job.
