Enabling 3D Applications in Public Cloud
The fast evolution of computing technologies and global deployment of data centers have made cloud computing a ubiquitous computing infrastructure. As a result, a large number of organizations begin to adopt cloud computing. With the growing popularity of cloud gaming and cloud virtual reality (VR), interactive 3D applications are becoming a major type of workload for the cloud. However, existing cloud systems are typically designed for traditional applications instead of 3D workloads. Therefore, there is a great need to optimize the performance and system efﬁciency of cloud 3D systems, to meet the Quality-of-Service (QoS) requirements, as well as the sustainability of data centers. More speciﬁcally, several critical research gaps are identiﬁed that hinder the migration of 3D applications to public cloud.
The ﬁrst research gap is lack of open and reliable research infrastructure, including benchmarks and performance analysis tools. Building a cloud 3D benchmarking framework is non-trivial. The challenges come fromgenerating human-like inputs under various system/application non-determinism and dissecting the performance of complex graphics systems. To solve these fundamental problems, we present the design of novel research infrastructure, Pictor, for cloud 3D applications and systems. Pictor employs AI to mimic human interactions with complex 3D applications. It can also track the processing of user inputs to provide in-depth performance measurements for the complex software and hardware stack used for cloud 3D graphics rendering. With Pictor, we also designed a benchmark suite with six interactive 3D applications. Performance analyses were conducted with these benchmarks to characterize 3D applications in the cloud and reveal new performance bottlenecks. To demonstrate the effectiveness of Pictor, we also implemented two optimizations to address two performance bottlenecks discovered in a state-of-the-art cloud 3D graphics rendering system, which improved the frame rate by 57.7% on average.
The second research gap is system efﬁciency for cloud 3D, including resource, memory, and energy efﬁciency. Current cloud 3D systems experience high resource and energy inefﬁciency due to excessive frame rendering. Therefore, the regulation of frame rates is required to improve energy and resource efﬁciency by reducing the cloud-client FPS gap and excessive rendering. However, due to the high variation in frame processing time and the use of rendering delays, existing FPS regulation solutions cause low client FPS and long motion-to-photon (MtP) latency, resulting in violations of QoS goals. We designed a novel cloud 3D FPS regulation solution, called OnDemand Rendering (ODR), which can effectively reduce the FPS gap while maintaining high frame rate and low motion-to-photon (MtP) latency. ODR employs multi-buffering, dynamic rendering delay/acceleration, and input processing prioritization to reduce excessive rendering, and ensure QoS satisfaction in terms of both frame rate and latency. These results also showed that by improving QoS, ODR makes it feasible to deploy 3D applications to current public clouds.
The third research gap is high network bandwidth usage, because of the limited compression rate of traditional handcrafted image/video compression algorithms. Traditional image/video compression methods have low compression rates and high network bandwidth usage, making it challenging to stream large amount of visual data on wide area networks (WAN) forsome cloud 3D applications. Recently, AI-based image compression has shown great potential on further increasing the image compression rate, but the computing time of which is very high. This work focuses on reducing the computing time of learned image compression to make it one step closer to the real-time requirement of cloud gaming and VR. We proposed application-speciﬁc compression to reduce the model complexity to speedup model computation time. Evaluations show that our approach has signiﬁcantly accelerated image compression/decompression without harming the image quality, making learned image compression potentially viable for cloud 3D.
Overall, this dissertation (1) deepens the understanding of the characteristics of 3D applications in the public cloud and, (2) develops new schemes to improve the performance, QoS, and resource efﬁciency of cloud 3D systems at low cost. (3) explores promising techniques related to image compression. These optimizations and works further enable 3D applications in public cloud.