This preview shows pages 1–2. Sign up to view the full content.
This preview has intentionally blurred sections. Sign up to view the full version.View Full Document
Unformatted text preview: Structure and Performance of the Direct Access File System Kostas Magoutis, Salimah Addetia, Alexandra Fedorova, Margo I. Seltzer Division of Engineering and Applied Sciences, Harvard University Jeffrey S. Chase, Andrew J. Gallatin, Richard Kisley, Rajiv G. Wickremesinghe Department of Computer Science, Duke University Eran Gabber Lucent Technologies - Bell Laboratories Abstract The Direct Access File System (DAFS) is an emerg- ing industrial standard for network-attached stor- age. DAFS takes advantage of new user-level net- work interface standards. This enables a user-level file system structure in which client-side function- ality for remote data access resides in a library rather than in the kernel. This structure addresses longstanding performance problems stemming from weak integration of buffering layers in the network transport, kernel-based file systems and applica- tions. The benefits of this architecture include lightweight, portable and asynchronous access to network storage and improved application control over data movement, caching and prefetching. This paper explores the fundamental perfor- mance characteristics of a user-level file system structure based on DAFS. It presents experimental results from an open-source DAFS prototype and compares its performance to a kernel-based NFS implementation optimized for zero-copy data trans- fer. The results show that both systems can deliver file access throughput in excess of 100 MB/s, sat- urating network links with similar raw bandwidth. Lower client overhead in the DAFS configuration can improve application performance by up to 40% over optimized NFS when application processing and I/O demands are well-balanced. 1 Introduction The performance of high-speed network stor- age systems is often limited by client overhead, such as memory copying, network access costs and protocol overhead [2, 8, 20, 29]. A related source of inefficiency stems from poor integration of ap- plications and file system services; lack of con- trol over kernel policies leads to problems such as double caching, false prefetching and poor concur- rency management . As a result, databases and other performance-critical applications often bypass file systems in favor of raw block storage access. This sacrifices the benefits of the file system model, including ease of administration and safe sharing of resources and data. These problems have also motivated the design of radical operating system structures to allow application control over resource management [21, 31]. The recent emergence of commercial direct- access transport networks creates an opportunity to address these issues without changing operating sys- tems in common use. These networks incorporate two defining features: user-level networking and re- mote direct memory access (RDMA). User-level net- working allows safe network communication directly from user-mode applications, removing the kernel from the critical I/O path. RDMA allows the net- work adapter to reduce copy overhead by accessing...
View Full Document
This note was uploaded on 11/12/2011 for the course CE 726 taught by Professor Staf during the Spring '11 term at SUNY Buffalo.
- Spring '11