Network infrastructure support for cloud services: Measurement and evaluation

This dissertation aims to explore measurement and evaluation methodologies for network infrastructure support in cloud services, in particular, leveraging existing Domain Name System (DNS) and Content Distribution Network (CDN) infrastructure to measure and evaluate network performance. This work is based on four major case studies: measuring and evaluating CDNs, estimating packet loss rate between arbitrary end-hosts, evaluating network performance for cloud services, and examining benefits of split-TCP proxies for cloud services. These cases are representatives of critical parts of today's Internet, and can make an impact on production system.;In the study of CDNs, we conduct extensive measurements that accurately characterize the performance of two large-scale leading commercial CDNs with radically different design philosophies: one enters deep into ISPs, while the other brings ISPs to home. Our measurement techniques can be adopted by CDN customers to independently evaluate the performance of CDN vendors. They can also be used by a new CDN entrant to choose an appropriate CDN design and to locate its servers. Furthermore, we quantify the potential gains of hybrid CDN-P2P for these two CDN companies by considering ISP-friendly P2P distribution schemes which work in conjunction with the CDNs to localize traffic within regions of ISPs. We use two recent, real-world traces -- a video-on-demand and a large-scale software update trace to demonstrate the benefits.;Estimate of packet loss rate between arbitrary Internet hosts is critical for many large-scale distributed applications, including overlay routing, P2P media streaming, VoIP, and edge-server location in CDNs. In this dissertation, we present Queen, a new methodology for estimating packet loss rates between arbitrary end-hosts by finding two DNS servers near them and determining the packet loss rates between these two DNS servers. Queen requires neither additional infrastructure deployment nor control of end-hosts.;To optimize network performance, cloud service providers can co-locate production servers in well-connected Internet eXchange (IX) points, deploy data centers in additional locations, or contract with external CDNs. Some options can be very costly, and some may not improve performance significantly. In the investigation of cloud services, we develop a measurement system, AdMeasure, which not only evaluates the performance of current architecture, but also accurately predicts the performance improvement of major infrastructure changes. AdMeasure uses active content transparently loaded into end-user systems, and performs latency and throughput measurements between the end systems and cloud service entities. To assist cloud services evaluating hypothetical deployments, we propose new methodologies for latency and throughput measurements. We apply these methodologies and perform a large-scale latency and throughput study of the Internet. Cloud service providers would be able to estimate a prior performance gain of the various options before sinking significant capital expenditures into major infrastructure changes.;Last but not least, we examine the benefit of split-TCP proxies deployed in an operational world-wide network for accelerating cloud services. We consider a fraction of a network consisting of a large number of satellite datacenters, which host split-TCP proxies, and a smaller number of mega datacenters, which ultimately perform computation or provide storage. Using web search as an exemplary case study, we conduct detailed measurements to characterize how individual components, including proxy stacks, network protocols, packet losses and network load, can impact the latency.