Running virtual machines on network file systems (NFS) provides a number of architectural advantages, starting with the fact that virtual machines use files to store their image information. Since network-attached storage (NAS) systems are built from the ground up for file management, the administrative time and resources to oversee a large number of virtual machine files is less than what is required to manually assign and place those files on individual logical unit numbers (LUNs).

New offerings, such as scalable caching appliances, now allow customers to extract more performance out of their existing NAS and NFS infrastructure. In fact, there are environments where NFS can perform as fast, if not faster, than other storage solutions based on Fibre Channel or iSCSI. By deploying NAS and NFS storage solutions, administrators can forgo many of the complexities associated with maintaining separate Fibre Channel adapters, switches and storage systems.

The first aspect of improving performance with NFS is the ability to remove the need for a host-based file system.

The necessity of Fibre Channel adapters is also one of the biggest problems in virtual environments, since rack-mounted servers often have a limited number of slots available for adapter cards. An Ethernet and IP-based solution avoids excessive adapters and can make effective use of embedded Gigabit Ethernet ports on the mother- board.

When provisioning virtual disks with NFS, new data stores can be added quickly by mounting a file system. There is no need to create or provision LUNs, as is the case with Fibre Channel or iSCSI data stores. In addition, some NFS systems allow for expanding and decreasing NFS volumes on the fly, and incorporate thin provisioning features by default, saving disk space. Users also can provision new virtual disks from an NFS snapshot, allowing rapid deployment of new virtual disks from a copy/snapshot of an existing virtual disk.

Managing virtual disks is typically easier with the file-management tools of NAS systems compared to manual handling of files within a LUN. Managing data stores between virtual servers is also simpler with NFS. With Fibre Channel or iSCSI, sharing data stores requires manually matching LUN IDs.

Protecting and recovering virtual disks can become a significant task for administrators. With a limited number of virtual machines, managing snapshots with host-based tools is practical and reasonable, but as the number of virtual machines increases, the need for more powerful tools escalates. Typically, the most powerful snapshot tools are based at the storage subsystem because the snapshot processing is close to the data and does not require traversing the network.

Because NFS systems are designed specifically for file-centric environments, they can be useful for file-level snapshot-and-restore capabilities. For example, they typically include the option to handle single file restore compared to volume-level restore capabilities of block-based approaches. The latter requires significantly more overhead in terms of time and storage resources for file-level recovery.

The first aspect of improving performance with NFS is the ability to remove the need for a host-based file system. Removing the host-based file system container and offloading that to the storage system clears the bottlenecks of a host-based file system. This sets the stage for separating functions between the host and storage system, such as snapshots.

Gary Orenstein

Another critical aspect of performance with virtualization is the maximum number of outstanding I/Os allowed for LUNs. With block-based approaches, this maximum number can artificially cap performance, because it must be shared across all of the virtual machines using the LUN. While the maximum LUN queue depth varies by host bus adapter and manufacturer, enabling the maximum number of outstanding I/Os (e.g., 256) can lead to considerable latency.

With NFS, there is no such I/O limitation. Consequently, increasing the bandwidth and connectivity between the virtual machines and the data store will continue to improve I/O performance.

Gary Orenstein is the vice president of marketing at Gear6. He is also the author of "IP Storage Networking: Straight to the Core," a book that outlines the business value of enterprise storage technology.

For more information (click here)