SIGGRAPH 2016 Post Mortem


by Ethan Estrada
Tags: SIGGRAPH, CGI, MetaPipe

I had the opportunity to attend SIGGRAPH this year as a representative of MetaPipe. For as fun as it was to meet with representatives of various software and cloud companies, I am happy to be able to get my nose back to the grindstone on MetaPipe. I’m extremely excited about the tech we are working on for the service and I can’t wait to get it to the point where we can start demoing it publicly to get everyone’s feedback. We started a lot of converstaions at SIGGRAPH with various companies that we feel fairly confident will turn into more concrete agreements in the near future.

I got an “Exhibits Only” pass during the convention; we are trying to be as scrappy, lean, and frugal as we can be so splurging on something like a “Full Conference” pass would definitly be overkill especially when Avere provided us with a code for a free “Exhibits Only” pass. Besides, all of the people we wanted to talk to were on the exhibit floor anyway. Maybe in future years we will get more complete passes to check out presentations to see the state of the industry, but for now we are keeping it simple. Ideally we will be able to justify having our own booth next year!

Even though I’ve heard SIGGRAPH has shrunk in attendence some in recent years, it is still very impressive in size and scope. I think this in part is from the decreased focus on the job fair and a greater focus on interconnect between vendors. SIGGRAPH seems less useful for individual artists looking for work than it used to be.

The biggest focus at the convention seemed to be VR and AR. Pretty much every booth from software vendors featured VR workflows. The new VR camera from Houdini, CaraVR addon for Nuke from The Foundry, the (not yet publicly available) VR toolset for Mocha from Imagineer systems, and more from other companies as well.

Not surprisingly, cloud was also a big focus (which was a big attraction to us). Many softwares packages such as deadline from Thinkbox software and Tractor from Pixar are starting to have more and more cloud based features, such as extremely granular billing for various software packages (from Thinkbox) and intelligent autoscaling features (from Tractor, and if I remember correctly, Thinkbox was working on similar stuff as well).

Autoscaling is an interesting problem in and of itself; scale too quickly and users will rack up charges for compute they didn’t really use efficiently. Scale too slowly and what should only take a few minutes could end up taking hours. Some workloads are heavy per frame and some are light. When do you scale up? How automagical can you make it for the user so that they have fewer knobs to deal with? Can you base it off of the basic priority parameter that most render schedulers already have, or do you need a different mechanism for it? Perhaps something like the expected average render time per frame (or task)? What about distributed simulations that are parallel, but need to be in lockstep? They should probably all spin up at the same time, but how do you signal this to the scheduler in a relatively turnkey way?

During the convention, I either asked these questions to the various developers on the farm manager teams, or thought of them after the fact due to mulling over the questions I had asked or information I had learned. It isn’t an easy problem to solve and one that I am sure will have many more years of time and effort poured into it before well rounded solutions come to market.

The post SIGGRAPH 2016 Post Mortem appeared first on Mister Fidget.