By Ann Silverthorn
—Panasas today announced the planned release this summer of key components of its proprietary DirectFLOW client software for Linux to the open source community. The motivation for the release is to accelerate adoption of the parallel network file system (pNFS) standard, which addresses storage I/O bottlenecks, and to spur end-user deployment of parallel storage solutions.
Panasas develops high-performance storage systems designed to help companies maximize the benefits of Linux clusters by eliminating the storage bottleneck created by older network-storage technologies. With the company's out-of-band DirectFlow data path, all Linux cluster nodes are empowered with a small installable file system called the DirectFLOW client. This client enables direct communication between Linux servers and Panasas StorageBlades.
"Changes on the computing side in high performance computing [HPC] have been all about clusters," says Larry Jones, vice president of marketing for Panasas. "Multi-core processors and applications take advantage of all the cores and the parallelism that's available."
Jones says these changes have opened the way for storage architecture shifts. "NFS was originally designed as a serial protocol—one lane in and one lane out," he says. "Now that you have multiple lanes, even down at the processor level, it looks like California-style freeways in the data center. We're looking to evolve the storage systems to achieve the same level of parallelism that's going on in the compute side."
It was Garth Gibson, Panasas founder and chief technology officer, who authored the pNFS problem statement to the IETF in 2004. Panasas has been providing technical input to the pNFS specification since that time.
The IETF NFSv4 subcommittee is expected to conclude its work on the pNFS protocol as part of the release of NFS version 4.1 request for comment (RFC) later this year. The ratified pNFS RFC is expected this fall. Parallel NFS enables direct parallel data transfer between clients and storage devices, without the need for expensive filer heads, and up to a 40% increase in speed. Jones says the storage can be block-based Fibre Channel, file-based NFS servers, or object servers.
"NFS version 4.1 provides the capability for parallel access to storage," says Jones. "In a typical NFS environment all the traffic goes through a server. pNFS pulls the server off to the side and lets the clients go directly and in parallel to the storage devices. So the servers only manage the metadata instead of also having to retrieve the data from the storage and hand it to the client. This opens up that parallel freeway from the client straight through to storage."
Jones says the pNFS standard minimizes risk. "Having a standard that validates interoperability is important from the users' perspective because it keeps them from being locked into a single vendor," he says. "The other advantage for users is that they get improved manageability by removing all the servers from the I/O path. It eliminates the load-balancing that typically takes place between the clients and the servers."
Joining Panasas in the consortium to advance pNFS are IBM, EMC, Network Appliance, Sun Microsystems, and the University of Michigan. Support is expected for Linux, Windows, and Unix versions such as Solaris and AIX.
The pNFS architecture is based in part on the Panasas DirectFLOW protocol. Panasas is releasing the source code of the DirectFLOW client for Linux to developers, specifically the storage access manager, OSDclient, Object iSCSI, and other network layers and parts of the Panasas libraries. It will be available to the storage community on the Panasas Website (www.panasas.com) and at www.pnfs.com.