This Computer Systems Research (CSR) Core Area project focuses on the grand challenge of achieving low end-to-end latency (under 100 ms) for web applications. Sub-100-ms latencies makes web applications feel as responsive as local applications that run on desktops/mobiles, which has enormous implications for enhanced user experiences, faster adoption of cloud computing, and for e-commerce revenues. Achieving such low latency is especially challenging because (a) todays web download process is a poor fit to high-latency cellular networks, and (b) the trend in web applications toward more personalized, dynamic content which makes it difficult to cache mutable web content close to users; the storage is typically held in distant centralized data-centers (DCs). Much of the previous work in caching dynamic content offers only weak consistency guarantees (typically, eventual consistency), but many interactive scenarios require the stronger causal model. The project will result in the design and evaluation of SmartEdge which offers low latency for mobile web applications while guaranteeing causal consistency for mutable backend storage.
This project addresses the above challenges via a comprehensive approach called SmartEdge which moves both the application functionality and the back-end storage to the network edge. To address the latency gap between cell devices and desktops, this research employs redundant execution on the cloud and proactively pushing data to the client. SmartEdge leverages the fact that redundant execution may be approximate in the context of web applications to drastically reduce the overheads of redundant computation, which enables the edge to scale to hundreds of thousands of clients. To address the challenge of scalable, consistent backend storage, SmartEdge (a) only partially replicates data in a subset of DCs (unlike existing techniques which require full replication of all data in all the DCs), and (b) supports large-scale and hierarchical edge caching (unlike existing geo-replicated storage systems that are limited to a small number of DCs). These techniques not only enable latency reduction by placing data closer to users but also achieve an order of magnitude lower replication cost. In addition, the project will extensively involve Ph.D, Masters and undergraduate students in research activities, and results from the projects will be published in premier scientific conferences. To support educational activities related to cloud computing, a "data-center-in-a-box" kit will be released.
