The Slice project explores techniques for building unified massive-data storage systems from inexpensive components connected by a high-speed network. Here are some of the project's goals and premises:

• Storage at network speed. With the emergence of Gigabit Ethernet as a commodity, the network will become the fastest path to I/O for most systems on Ethernet networks (LANs). A key technical challenge is to manage disks and memories across the network as a unified resource whose performance tracks rapid advances in network technology rather than the slower improvements in disk speeds.
• Internet/LANs as a scalable storage backplane. Commercial systems increasingly provide shared storage by interconnecting storage devices and servers with dedicated Storage Area Networks (SANs), e.g., FibreChannel. Yet recent order-of-magnitude improvements in LAN performance have narrowed the bandwidth gap between SANs and LANs, undermining the need for a separate SAN. LAN-based network storage systems can approach SAN systems in terms of incremental scalability, reliability, ease of administration, and performance, while offering high-speed storage access to standard LAN-attached clients at lower cost.
• Decentralized file service structure. Client/server LAN file services were dominant through the late 1980s and 1990s and up to today. Most file system functions are handled by central file servers exporting one or more storage volumes through an Internet-based network file system protocol (e.g., NFS). The latest generation of network file servers are often confusingly referred to as network-attached storage (NAS) appliances, in order to contrast with the SAN approach. To deliver on the potential of high-speed LANs, NAS systems must evolve away from the client/server model toward the more decentralized service structure of the SAN systems, distributing storage functions across a fluid collection of cooperating servers and storage nodes.
• Intelligent block placement and movement. Overall storage system performance and reliability are determined primarily by the policies for distributing data blocks across storage nodes, placing blocks on disks, and timing data movement between slow disks and fast memory. We are experimenting with policies and mechanisms for informed prefetching, network memory caching, mirrored striping, application-directed block placement, and storage management that adapts to available resources.

The Slice file service is implemented as a set of loadable kernel modules in the FreeBSD operating system. A Slice is configured as a combination of server modules that handle specific file system functions: directory management, raw block storage, efficient storage of small files, and network caching. Servers may be added as needed to scale different components of the request stream independently, to handle a range of data-intensive and metadata-intensive workloads. Slice is designed to be compatible with standard NFS clients, using an interposed network-level packet translator to mediate between each client and an array of servers presenting a unified file system view.

• Duke Cluster Lab
• Trapeze project
• Publications
• Fstress network file system benchmark

Personnel

• Faculty
  • Jeff Chase
  • Amin Vahdat
• Students
  • Darrell Anderson
  • Richard Kisley
  • Ken Yocum
• Research Staff
  • Drew Gallatin

Funding