TECHNOLOGY

Scaling File Systems to the Next Level

File systems have traditionally been "monolithic", with a single server managing the metadata for the entire file system. This simple approach has reached its limits and is unable to meet the performance, availability and manageability requirements being demanded by today's information-hungry applications. The next generation of file systems must distribute the metadata management function across multiple servers allowing parallel access from these servers so as to deliver the levels of performance required. Additionally, metadata must be managed in a modular fashion that allows construction of very large namespaces that are not only tolerant but also resilient to faults in that the entire namespace does not become unavailable if some portion of it is inaccessible. The architecture must allow scaling of capacity and performance independently of each other, a user should be able to add more capacity to the file system without adding more servers and vice versa.

The Need for a Parallel File System

Breaking the single-server barrier of file system metadata management requires the construction of a Parallel File System (alternatively referred to in the industry as "distributed" or "global" file system) whereby multiple servers can concurrently access and modify the same file system. To date, two approaches have been used to try to solve this very difficult problem. The hierarchical approach distributes metadata management responsibility among servers based on the tree structure in the file system. Put simply, each server is given ownership of an entire sub-tree of the file system. Since the structure of the directory tree and location of files within it are determined by users and their applications and not by the file system and as such are unpredictable, this approach is prone to "hot spots" servers that end up owning much of the commonly accessed data and thus become performance bottlenecks. The second approach is a more symmetric one where metadata ownership is determined dynamically at run time, at the level of each individual file. When a server wishes to access a file, it must first determine which server currently has metadata ownership for that file, request and receive the ownership from that server, and then proceed with I/O. The fine grained nature of the distributed metadata management in this architecture leads to enormous overheads in the general case, making this approach unsuitable for all but the most embarrassingly parallel applications.

IBRIX Fusion Segmented File System

IBRIX has developed a new approach, a unique architecture called a Segmented File System (U.S. Patent 6,782,389). In traditional parallel computing terms, this architecture may be described as a loosely-coupled approach to distributing metadata. Essentially, the segmented architecture is based on the "divide and conquer" principle similar to the telephone networks in the United States, wherein each switch only keeps detailed information about the area codes it services and simply routes calls to other area codes to the appropriate switches. As such, metadata propagation and reconciliation no longer create a bottleneck, resulting in linear scalability. The segmentation also provides fault containment and modularity which makes the namespace fault resilient and easy to manage and grow.

To learn more about IBRIX's patented Segmented File System architecture, please visit our Literature page to download product literature.