Large-scale network services can consist of tens of thousands of machines running thousands of unique software configurations spread across hundreds of physical networks. Testing such services for complex performance problems and configuration errors remains a difficult problem. Existing testing techniques, such as simulation or running smaller instances of a service, have limitations in predicting overall service behavior at such scales. Testing large services should ideally be done at the same scale and configuration as the target deployment, which can be technically and economically infeasible. We present DieCast , an approach to scaling network services in which we multiplex all of the nodes in a given service configuration as virtual machines across a much smaller number of physical machines in a test harness. We show how to accurately scale CPU, network, and disk to provide the illusion that each VM matches a machine in the original service in terms of both available computing resources and communication behavior. We present the architecture and evaluation of a system we built to support such experimentation and discuss its limitations. We show that for a variety of services---including a commercial high-performance cluster-based file system---and resource utilization levels, DieCast matches the behavior of the original service while using a fraction of the physical resources.

Modern day datacenters host hundreds of thousands of servers that coordinate tasks in order to deliver highly available cloud comput-ing services. These servers consist of multiple hard disks, memory modules, network cards, processors etc., each of which while care-fully engineered are capable of failing. While the probability of seeing any such failure in the lifetime (typically 3-5 years in indus-try) of a server can be somewhat small, these numbers get magni-fied across all devices hosted in a datacenter. At such a large scale, hardware component failure is the norm rather than an exception. Hardware failure can lead to a degradation in performance to end-users and can result in losses to the business. A sound under-standing of the numbers as well as the causes behind these failures helps improve operational experience by not only allowing us to be better equipped to tolerate failures but also to bring down the hardware cost through engineering, directly leading to a saving for the company. To the best of our knowledge, this paper is the first attempt to study server failures and hardware repairs for large data-centers. We present a detailed analysis of failure characteristics as well as a preliminary analysis on failure predictors. We hope that the results presented in this paper will serve as motivation to foster further research in this area.

ModelNet is a network emulator designed for repeatable, large-scale experimentation with real networked systems. This talk introduces the ideas behind ModelNet that have made it a successful experimental platform. Beyond these core concepts, the talk highlights the latest additions to our methodology to test the next generation of network protocols and applications. Many of these developments address the datacenter compute environment: high-capacity networks, sophisticated infrastructure software (storage and virtualization), and complex network load. While these efforts significantly extend ModelNet's capabilities, there remain a number of open challenges, including incorporating new performance objectives (energy) and multicore architectures.

Today's cloud-based services integrate globally distributed resources into seamless computing platforms. Provisioning and accounting for the resource usage of these Internet-scale applications presents a challenging technical problem. This paper presents the design and implementation of distributed rate limiters, which work together to enforce a global rate limit across traffic aggregates at multiple sites, enabling the coordinated policing of a cloud-based service's network traffic. Our abstraction not only enforces a global limit, but also ensures that congestion-responsive transport-layer flows behave as if they traversed a single, shared limiter. We present two designs - one general purpose, and one optimized for TCP - that allow service operators to explicitly trade off between communication costs and system accuracy, efficiency, and scalability. Both designs are capable of rate limiting thousands of flows with negligible overhead (less than 3% in the tested configuration). We demonstrate that our TCP-centric design is scalable to hundreds of nodes while robust to both loss and communication delay, making it practical for deployment in nationwide service providers.

This paper presents Swing, a closed-loop, network-responsive traffic generator that accurately captures the packet interactions of a range of applications using a simple structural model. Starting from observed traffic at a single point in the network, Swing automatically extracts distributions for user, application, and network behavior. It then generates live traffic corresponding to the underlying models in a network emulation environment running commodity network protocol stacks. We find that the generated traces are statistically similar to the original traces. Further, to the best of our knowledge, we are the first to reproduce burstiness in traffic across a range of timescales using a model applicable to a variety of network settings. An initial sensitivity analysis reveals the importance of capturing and recreating user, application, and network characteristics to accurately reproduce such burstiness. Finally, we explore Swing's ability to vary user characteristics, application properties, and wide-area network conditions to project traffic characteristics into alternate scenarios.

In earlier work we proposed a way for a Web server to detect connectivity information about clients accessing it in order to take tailored actions for a client request. This paper describes the design, implementation, and evaluation of such a working system. A Web site has a strong incentive to reduce the `time-to-glass' to retain users who may otherwise lose interest and leave the site. We have performed a measurement study from multiple client sites around the world with various levels of connectivity to the Internet communicating with modified Apache Web servers under our control. The results show that clients can be classified in a correct and stable manner and that user-perceived latency can be reduced via tailored actions. Our measurements show that classification and determination of server actions are done without significant overhead on the Web server. We explore a variety of modified actions ranging from selecting a lower quality version of the resource to altering the manner of content delivery. By studying numerous performance related factors in a single unified framework and examining both individual actions as well as combination of actions, our modified Web server implementation shows the efficacy of various server actions.