Local renders faster than farm renders.

[aabrams]

Question:
My renders on the farm are taking unusually long. Renders that take 6-14 minutes per frame rendered locally are taking on average 40-80 minutes or more on the farm. I have tried deleting all extraneous objects and render layers from scene file before submitting and it hasn't seemed to have an effect. The renders being produced are fine, I am just wondering why they are taking so long on the farm. Also, I have noticed that the dedicated render nodes are taking 15-20 minutes longer than the regular Montgomery machines, which seems to be contrary to what I have heard about the speed comparison between the dedicated nodes and "monty" machines.


Answer:
In some cases, local renders can use up to 8 processors at a time, while the nodes on the render farm will not use all its CPU cores (If it's an 8 Processors machine, it will use about 3-4 of them on the farm), because the nodes are gathering render stats and transferring files through the network simultaneously.
In my experience, it's acceptable that the render time of one frame on render farm is about 4 times that of a frame rendered locally, especially on the computers with 8 processor cores. However, the total render time will still be significantly less than your local render time if a project has hundreds of frames to render.

(Answer submitted by Ling Li)

[aabrams]

I believe Ling's explanation is quite good. Putting one frame on the renderfarm, has always been a trivial idea in my eyes. Besides the element of time it takes to copy and set up a job on the farm, our farm does not work on a "bucket to bucket" system. One render node (computer) will render one frame (.jpg, .tif etc.) until it's done. The farm will not take a single frame, and parse that frame out to all the nodes available, to render pixel buckets. There are some utilities that will do that, in fact Mental Ray Satellite will do this as well as Houdini's Render Queue. Our farm currently is not set up that way. It does make sense knowing that the computers in 221 are the fastest computers (hence the rooms popularity), and when looking at the rendertimes throughout frame logs during renderfarm jobs, 221 computers statistically are faster. By default, the first frames of your renders always go to the render nodes dedicated in the systems office. They are fast, but I believe the 221 workstations will be faster in terms of multi-threading, core usage, and ram count.

Although not stressed in many cases, I believe backface culling and double sided objects will nominally impact your render times. I had a project in Houdini which involved a liquid object, that was being viewed top down. It had an environment map, global-illumination and a pretty robust liquid shader with displacements. On local renders, frames were averaging 2-5 minutes, depending on motion complexity. When I submitted about 150 frames to the farm, my render times jumped to 90 minutes a frame. What I figured out was, when rendering locally, the ground plane I had as the liquid objects bottom served as a cutoff for reflections and refractions from the under side of the liquid object. When on the farm, without the viewport, the liquid object was being rendered from all angles, computing way more reflections and refractions than I needed. Once I turned on remove backface on the liquid object, all was well.

I'm no expert, but certainly everything has it's learning curve and surprises.

~T.J. Zoltner

[aabrams]

One more thing to keep in mind is final gather. If the project is set up under windows the farm will be unable to read the final gather map and therefore recompute it for every frame. This has a severe impact on rendertimes.
Student have to adjust the path for the fg map in the folder structure on the renderfarm manually to make this work. (As explained in detail in the forum under rendertimes).
Too many people are still not aware of this.

Sandro