Norman Hutchinson

The classical move from single-server applications to scalable cloud services is to split the application state along certain dimensions into smaller partitions independently absorbable by a separate server in terms of size and load. Maintaining data consistency in the face of operations that cross partition boundaries imposes unwanted complexity on the application. While for most applications many ideal partitioning schemes readily exist, First-Person Shooter (FPS) games and Relational Database Management Systems (RDBMS) are instances of applications whose state can’t be trivially partitioned. For any partitioning scheme there exists an FPS/RDBMS workload that results in frequent cross-partition operations. In this thesis we propose that it is possible and effective to provide unpartitionable applications with a generic communication infrastructure that enforces strong consistency of the application’s data to simplify cross-partition communications. Using this framework the application can use a sub-optimal partitioning mechanism without having to worry about crossing boundaries. We apply our thesis to take a head-on approach at scaling our target applications. We build three scalable systems with competitive performances, used for storing data in a key/value datastore, scaling fast-paced FPS games to epic sized battles consisting of hundreds of players, and a scalable full-SQL compliant database used for storing tens of millions of items.

View record

Emerging interactive multimedia applications, such as real-time visualizations, animations, on-line games, virtual reality, and video conferencing have low latency interactions and continuous high resource (e.g., CPU processing and network bandwidth) demands. The combination of latency sensitive interactions and high resource demands is challenging for best-effort platforms, such as the Internet, general-purpose operating systems and Web browsers because these platforms have no timing or resource guarantees and tend to favor high utilization. When demands exceed available resources, it is impossible to process all computations and data in a timely fashion resulting in diminished perceived quality (e.g., frame rate) and brittle real-time performance. The mismatch between application demands and available resources is observed to varying degrees in all resources including network, processing, and storage.To deal with the volatility and shortage of resources, we build upon and extend the Priority-Progress quality adaptation model. Our approach enables applications to scale demands (up or down) based on available resources and to utilize the limited resources in processing the computations and data with more influence over perceived quality. We develop enhancement layers to improve timeliness and guarantee more consistent quality using quality adaptation while maintaining the strengths of the existing best-effort transports and execution platforms. DOHA, our execution layer, extends the Priority-Progress CPU adaptation to work in games and across multiple execution threads. The modified game has better timing, higher perceived quality, and linearly scalable quality with a small number of cores. Our transport layer, Paceline, introduces low latency techniques over TCP and exposes Priority-Progress adaptation as an essential transport feature improving upon TCP's end-to-end latency while preserving its fairness and utilization.

View record