Render Cost Driven Caching for Object-Based Media

Dean, Paul and Ramdhany, Rajiv and Porter, Barry (2025) Render Cost Driven Caching for Object-Based Media. In: 2025 IEEE International Conference on Autonomic Computing and Self-Organizing Systems Companion (ACSOS-C) :. IEEE. ISBN 9798331502164

Text (IEEE-conference-template-062824)
IEEE-conference-template-062824.pdf - Accepted Version
Available under License Creative Commons Attribution.
Download (710kB)

Abstract

The caching of popular content in fast-access memory has long been a staple of performance optimisation. A common assumption of cache protocols is that a cache miss generates a roughly equal level of work for any item. In this paper, we introduce a novel caching problem in which this assumption does not hold, in the emerging domain of Object-Based Media (OBM). Using OBM, a media experience is segmented and decomposed into individual objects, layers, and assets. These elements are then dynamically composed, during playback, by each user according to their preferences. In this domain, cache misses while generating a personalised media rendition will have highly varying costs depending on how expensive it is to render that segment of media. We analyse the challenges of caching based on render cost variability for object-based media delivery within self-adaptive systems, and present our work on a novel approach to cost-based cache partitioning and policy selection to reduce the total render cost. We present experimental results for the challenge of determining element retention in a cache, and identify a key challenge in determining retention rates via frame rate, request pattern, and GPU utilisation, creating a cost value per media object dependent on both velocity and render complexity. We examine different cache compositions using a single policy for uniform and non-uniform memory per object, with comparison to a recently proposed multi-queue implementation, studying the impact of partition adjustment. Our results demonstrate that cost-informed cache partitions produce a lower render cost for synthetic workload traces, lowered further through partition adjustment, despite a higher miss rate.