A Holistic Approach to Lowering Latency in Geo-distributed Web Applications

User perceived end-to-end latency of web applications have a huge impact on the revenue for many businesses. The end-to-end latency of web applications is impacted by: (i) User to Application server (front-end) latency which includes downloading and parsing web pages, retrieving further objects requested by javascript executions; and (ii) Application and storage server(back-end) latency which includes retrieving meta-data required for an initial rendering, and subsequent content based on user actions. Improving the user-perceived performance of web applications is challenging, given their complex operating environments involving user-facing web servers, content distribution network (CDN) servers, multi-tiered application servers, and storage servers. Further, the application and storage servers are often deployed on multi-tenant cloud platforms that show high performance variability. While many novel approaches like SPDY and geo-replicated datastores have been developed to improve their performance, many of these solutions are specific to certain layers, and may have different impact on user-perceived performance. The primary goal of this thesis is to address the above challenges in a holistic manner, focusing specifically on improving the end-to-end latency of geo-distributed multi-tiered web applications. This thesis makes the following contributions: (i) First, it reduces user-facing latency by helping CDNs identify and map objects that are more critical for page-load latency to the faster CDN cache layers. Through controlled experiments on real-world web pages, we show the potential of our approach to reduce hundreds of milliseconds in latency without affecting overall CDN miss rates. (ii) Next, it reduces back-end latency by optimally adapting the datastore replication policies (including number and location of replicas) to the heterogeneity in workloads. We show the benefits of our replication models using real-world traces of Twitter, Wikipedia and Gowalla on a 8 datacenter Cassandra cluster deployed on EC2. (iii) Finally, it makes multi-tier applications resilient to the inherent performance variability in the cloud through fine-grained request redirection. We highlight the benefits of our approach by deploying three real-world applications on commercial cloud platforms