Supporting peripheral devices, once a motivation for creating operating systems, has received little attention from the research community. However, device access, whether for storage, networking, or graphics, is a major source of complexity, unreliability, and cost for modern operating systems.

In this talk, I will present two approaches to treating device access as a first-class topic. First, we investigated the reliability of device drivers in the presence of faulty devices and found there are many drivers that will crash or hang when a device fails. We address this problem with Carburizer, a code-manipulation tool and associated runtime that detects and repairs such bugs.

I will then discuss our work on storage devices that support direct access by applications. New storage-class memory (SCM) technologies, such as phase-change memory, promise user-level access to non-volatile storage through regular memory instructions. We built Mnemosyne, a simple interface for programming with persistent memory that addresses two challenges: how to create and manage such memory, and how to ensure consistency in the presence of failures.

EMC is a global leader in enabling businesses and service providers to transform their operations and deliver Information Technology as a service. Fundamental to this transformation is cloud computing. Through innovative products and services, EMC accelerates the journey to cloud computing, helping IT departments to store, manage, protect and analyze their most valuable asset information in a more agile, trusted and cost-efficient way.

In this talk, you will gain an understanding of the complex underlying computer science problems that EMC is attempting to solve. You will see an overview of two divisions: Backup and Recovery Division and Isilon Storage Division. The talk will be highly interactive - so bring your questions.

Abstract: We analyze the I/O behavior of iBench, a new collection of productivity and multimedia application workloads. Our analysis reveals a number of differences between iBench and typical file-system workload studies, including the complex organization of modern files, the lack of pure sequential access, the influence of underlying frameworks on I/O patterns, the widespread use of file synchronization and atomic operations, and the prevalence of threads. Our results have strong ramifications for the design of next generation local and cloud-based storage systems.

(This is a PRACTICE TALK for SOSP '11)

BIO: Tyler Harter is a 2nd-year graduate student at the University of Wisconsin-Madison where he works as a research assistant with Professors Andrea Arpaci-Dusseau and Remzi Arpaci-Dusseau in the Department of Computer Sciences. He received his B.S. in both Computer Science and Mathematics from the University of Wisconsin-Madison in 2010. Tyler is interested in how storage is used by home users and how file and storage systems can improve their experience.

Abstract: One of the key challenges in the data center today is the efficient use of data-center resources while providing services with service-level objectives (SLOs). The primary reason for the challenge is the dynamism in workload requirements over time. We propose dynamic instantiation of virtual caching appliances (caches as virtual machines) to handle the dynamism in workloads and thereby support storage SLOs efficiently. We have developed an SLO-based automation framework called Dynamite for cache instantiation, that includes low-overhead techniques to: determine the workloads that would benefit from caching, determine the appropriate cache size for these workloads, instantiate the cache and non-disruptively migrate the application, and finally warm the cache to quickly return to acceptable service levels. We have evaluated the effectiveness of the individual techniques using a variety of I/O traces and find that they are highly accurate despite approximations that significantly reduce monitoring overheads. And finally, the approach actually works! Using the complete pipeline on a case study involving interfering workloads shows that service levels can be met while utilizing resources efficiently.

*Bio:* Lakshmi N. Bairavasundaram is a member of technical staff in the Advanced Technology Group at NetApp. His research interests include storage systems, file systems, storage and data management, and fault tolerance. Lakshmi currently focuses on storage and data management abstractions and techniques. He joined NetApp after completing his Ph.D. in Computer Sciences (2008) at the University of Wisconsin-Madison under the supervision of Prof. Andrea Arpaci-Dusseau and Prof. Remzi Arpaci-Dusseau.