An Extended Evaluation of Two-Phase Scheduling Methods for Animation Rendering

Recently HP Labs engaged in a joint project with DreamWorks Animation to develop a Utility Rendering Service that was used to render part of the computer-animated feature film Shrek 2. In a companion paper [2] we formalized the problem of scheduling animation rendering jobs and demonstrated that the general problem is computationally intractable, as are severely restricted special cases. We presented a novel and efficient two-phase scheduling method and evaluated it both theoretically and via simulation using large and detailed traces collected in DreamWorks Animation's production environment. In this paper we describe the overall experience of the joint project and greatly expand our empirical evaluations of job scheduling strategies for improving scheduling performance. Our new results include a workload characterization of DreamWorks Animation animation rendering jobs. We furthermore present parameter sensitivity analyses based on simulations using randomly generated synthetic workloads. Whereas our previous theoretical results imply that worst-case performance can be far from optimal for certain workloads, our current empirical results demonstrate that our scheduling method achieves performance quite close to optimal for both real and synthetic workloads. We furthermore offer advice for tuning a parameter associated with our method. Finally, we report a surprising performance anomaly involving a workload parameter that our previous theoretical analysis identified as crucial to performance. Our results also shed light on performance tradeoffs surrounding task parallelization.