Distributed File Systems For Storage Area networks - Technology Information

Computer Technology Review, Jan, 2000 by Brad Kline

This article discusses some fundamental aspects of Distributed File Systems (DFS) for Storage Area Networks (SANs). Many desirable features make a SAN DFS attractive to a production environment that processes huge volumes of bulk data and requires high bandwith. Throughout the discussion, the traditional server-attached storage model is compared against the distributed network-attached storage model. The remainder of the article will focus on issues of DFS configuration, data placement strategies, cross platform data sharing, fault tolerance, and performance. Finally, we will show that distributed file systems can provide superior performance over traditional server-attached storage shared file systems.

Various shared file systems have been in use for more than fifteen years. They started as simple systems using file server protocols such as the Network File System (NFS). These products are still a very large part of today's shared storage solutions. With the advent of open standard technologies such as Fibre Channel and its ability to imbed SCSI commands, it is possible to connect mass storage directly into a network at relatively low cost. Using compatible Fibre Channel Host Based Adapters (HBAs), any work-station can address the storage and access data to and from it, just as it would do to a directly connected SCSI storage device. The network, which includes HBAs, hubs, switches, and network-attached storage, is called a Storage Area Network (SAN). Fibre Channel is a high-speed (100MB/sec) channel that is capable of connection distances measured in kilometers. The Fibre Channel protocol uses addressable nodes in such a way that storage devices can be configured in a network fabric rather than point-to-point. This capability provides a storage network that is highly scalable, has high performance, and allows complete sharing of any connected device.

The distinct difference between using the conventional server-attached storage shared file system and a SAN is that any workstation connected to the network fabric can directly access the network-attached storage devices. The new connectivity obviates the need for a server and promotes a more distributed approach to managing data. This concept has started a new surge of distributed storage product development and has been coined the "SAN solution."

Many vendors have committed resources to developing hardware products that comprise a SAN by producing a variety of Fibre Channel compatible hardware components. In and by themselves, they offer a high-speed access method to storage. What has been missing from the formula is the ability for software to leverage the distributed nature of the storage network. At the core of every operating system are file systems that abstract disk storage and allow multiple programs to reliably share this resource. Existing local file systems such as Microsoft's NTFS 1, SGI's XFS, and Apple's HFS assume that any visible storage, whether in a network or locally connected, is owned by the local workstation. If the file system accesses the storage without coordination with other workstations in the network and two or more inadvertently share access to the network-attached storage devices, data corruption occurs. The early SAN devices were used only in a server-attached shared storage environment because of this problem and due to no readily available software solution. In the last three or four years however, a number of mostly third-party vendors have begun developing and offering distributed file system solutions that exploit the SAN distributed environment.

This article will discuss how a SAN File System (SAN FS) operates. We will discuss the difference between a typical server-attached shared storage model using NFS and a distributed network-attached storage model using Mountaingate's CentraVision File System (CVFS). Then, we will address important SAN FS features such as configuration, data placement strategies, cross platform data sharing, fault tolerance, and performance.

STORAGE MODELS

* The Server-Attached Shared Storage Model. In a typical data processing center today, there may be many different methods of sharing storage among multiple workstations. The most prevalent storage model is server-attached. This model uses one or more large workstations or file servers to locally attach the shared disk storage. The storage is then shared over some network topology using protocols like NFS. The systems are inter-connected through network interfaces such as Ethernet, HiPPI, or ATM. Extensive effort has been made by numerous vendors to make the file server model a high performance solution. A problem exists in that the model must copy from the data storage to the server's memory before it can transmit the information to the requestor. The performance, therefore, is gated by the speed of the server's memory and processing power (Fig 1).

* The Distributed Network-Attached Storage Model. By making storage a part of the network rather than a part of the local server, a DFS can use a different access method. The storage model now changes from server-based to peer-to-peer between the workstation and the storage. The advantage to this model is that with proper system software the data can be transmitted directly from the storage to any application without using a file server as a middleman (Fig 2).