Collaborative Research: CNS Core: Medium: A Unified Prefetch Framework for Approximation-Tolerant Interactive Applications

交互性是许多面向用户的应用程序的核心需求，包括数据可视化、网络搜索和游戏。这些面向用户的交互式应用必须实现低延迟响应，以满足用户的需求，而在做出反应之前等待用户的决定并不总是能够满足这一点。另一种方法是预取数据，以预测用户的选择。预取的现有使用有几个限制：(1)它通常以一种特殊的方式为每个应用程序开发，并且不考虑所有优化方面；(2)它们没有明确地利用许多交互式应用程序的近似容受性。近似容忍意味着用户更喜欢快速但近似的结果，而不是完全正确但缓慢的结果。该项目设计了一个称为GPF的通用预取框架，它显式地将预测和调度与客户端应用程序解耦。可配置的预测模型估计未来不同时间间隔内请求的可能性，通用调度器使用这些预测来决定将哪些请求发送到客户机。这个框架在几个方面是新颖的：(1)客户端不是显式请求，而是偶尔向调度器提供预测，调度器在向客户端推送结果时考虑网络和资源条件；(2)GPF利用应用程序容差为大量候选请求发送部分结果，而不是为少数请求发送完整结果；(3)GPF根据延迟、网络和资源条件动态地改变预测器和调度器计算在客户端或服务器上的位置。支持性研究将来自网络社区的性能和调度思想与来自数据库和可视化社区的优化、存储和交互思想结合在一起。GPF将集成并消除用户感知的多个领域应用程序的延迟，包括数据可视化、媒体播放器、网页导航、车辆控制和游戏。多学科研究（网络、信息可视化和数据库系统）将整合到数据科学、数据库、网络和可视化课程中。软件将是开源的，并将对交互式应用程序的开发方式产生重大的、长期的影响。研究和教育材料的成果将通过研讨会、出版物和开源存储库传播。这些教育和推广计划将进一步增加对这一多学科主题的参与，这将导致大数据可视化技术、网络调度和优先级设计的持续进步，并最终使越来越多依赖或需要交互式应用程序来做出时间关键决策和发现的领域受益。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Interactivity is a core requirement for a wide range of user-facing applications, including data visualizations, web search and games. These user-facing, interactive apps must achieve low latency responses in order to satisfy users, which cannot always be met by waiting for a user's decision before reacting. An alternative is to pre-fetch data in anticipation of users' choices. There are several limitations to existing uses of prefetching: (1) it is often developed in an adhoc way for each application, and does not consider all optimization aspects, and (2) they do not explicitly take advantage of the approximation tolerant nature of many interactive applications. Approximation tolerance means that users prefer fast but approximated, over fully correct but slow, results.This project designs a General Prefetching Framework called GPF that explicitly decouples prediction and scheduling from the client application. A configurable prediction model estimates the likelihood of requests at different future time intervals, and a general scheduler uses these predictions to decide which requests to send to the client. This framework is novel in several ways: (1) rather than explicit requests, the client occasionally offers predictions to the scheduler, which considers network and resource conditions when pushing results to the client, (2) GPF exploits application tolerance to send partial results for a massive number of candidate requests, rather than full results for a few requests, and (3) GPF dynamically shifts placement of the predictor and scheduler computation on the client or server based on latency, network, and resource conditions.The supporting research brings together performance and scheduling ideas from the networking community with optimization, storage, and interaction ideas from the database and visualization communities. GPF will integrate and eliminate user-perceived application latency in applications across multiple domains, including data visualization, media players, webpage navigation, vehicular control, and games. The multidisciplinary research (networking, information visualization, and database systems) will be integrated into courses on data science, databases, networking, and visualization. Software will be open sourced, and will have significant, long-term impact on the way interactive applications are developed. The outcomes of the research and education material will be disseminated via workshops, publications, and open-source repositories. These education and outreach plans will further increase participation in this multidisciplinary topic that will lead to the continuing advancement of big data visualization techniques, network scheduling and prioritization designs, and ultimately benefit the increasing number of domains that rely on, or demand, interactive applications to make time critical decisions and discoveries.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.