作者: Mohammad Al-Fares , Khaled Elmeleegy , Benjamin Reed , Igor Gashinsky
关键词:
摘要: Fast-loading web pages are key for a positive user experience. Unfortunately, large number of users suffer from page load times many seconds, especially with embedded objects. Most this time is spent fetching the and its objects over Internet.This paper investigates impact optimizations that improve delivery content edge servers at Yahoo! Content Delivery Network (CDN) to end users. To end, we analyze packet traces 12.3M TCP connections originating across world terminating CDN. Using these traces, characterize user-connection metrics network, transport, application layers. We observe high Round Trip Times (RTTs) inflated round trips per download (RTT multipliers). Due inefficiencies in TCP's slow start HTTP protocol, found several opportunities reduce RTT multiplier, e.g. increasing Initial Congestion Window (ICW), using Appropriate Byte Counting (ABC), pipelining.Using live workloads, experimentally study micro effects on network connectivity, loss rate. evaluate macro overall time, use realistic synthetic workloads closed laboratory environment. find compounding pipelining ICW size can lead reduction by up 80%. also no one configuration fits all users, certain may help some while hurting others.