pNFS: Imagining the future and delivering on it.
Lots of buzz about the first pNFS Open Solaris code drop. For storage people the things that strike me are:
Unlimited Horizontal Scale: pNFS delivers a global namespace for file systems - think of the problem with NetApp, when you fill up one NetApp box, you get another, etc. and then you have to deal with remembering which files are on which box. With pNFS you get one file namespace, regardless of the number of servers.
It achieves this via an industry standard protocol (that's what the NFS v4.1 spec is defining, the over-the-wire protocol used between the clients needing to access a file, and the server that stores that file). This means that vendors can now start delivering interoperable, horizontally scalable solutions. In all other "clustered" file system solutions today, the implementation is proprietary – you have to get all the parts of the solution from one vendor.
The next interesting part is that we are making this available via open source, in order to continue to propagate the standard to ensure customers get high quality, interoperable, next gen solutions for file.
Bottom line: true seamless horizontal scale across multiple NAS boxes or file servers. A desktop application will no longer have to remember which NAS or server is storing their file ... imagine that!
Posted at 06:10PM Jul 16, 2007 by Nigel Dessau in Sun Storage | Comments[2]
Nigel, will Sun actually do something with pNFS (and the rest of NFS v4)? Sun does not have a good record in storage. STK gave Sun customers, and innovation and engineering in tape drives and robotics, but the big money in storage continues to be disk. And NFS v4 could usher in the biggest changes to disk storage since the advent of Fibre channel. NFSoRDMA offers filesystem flexibility at block performance levels. pNFS offers, as you point out, non-proprietary, heterogeneous clustered filesystems. pNFS metadata running on NFSoRDMA with block data transmitted by iSER, FC, or FCoE would offer heterogeneous clustered filesystems with SAN levels of performance (and would directly compete with, and likely succeed, QFS). A pNFSoRDMA using iSER on iWARP would offer a "single wire" solution for a clustered filesystem, which would greatly simplify implementation.
Every NFS server requires a computer with a native filesystem. WAFL has much to do with the success of NetApp.
Sun has the ultimate native filesystem, ZFS. It has a highly network performant OS in Solaris. Sun has its old SNDR product, which offers replication for a Solaris based fileserver. Sun even has the ability to server FC blocks out of a Solaris server with its old STE software.
Now if Sun can cluster a couple of x86 or Niagara servers together, connect them to an FC JBOD array, run RAID1, RAIDZ, and/or RAIDZ2 across them, and push out NFS, pNFS, NFSoRDMA, and pNFSoRDMA over 10Gb Ethernet, and fully appliantize it would be a heck of a storage box.
You will have to out NetApp NetApp. If you don't pull it off, get ready, because my guess is NetApp will be the big winner as more and more storage moves to Ethernet and NFS.
The question is, can Sun pull it off?
Posted by Mark on July 16, 2007 at 08:56 PM MDT #
Posted by Don Traub on July 17, 2007 at 09:51 AM MDT #