NFS client with AsyncIO and without use of buffer cache

[Walter Kießl]

Hi there,

Sun has implemented a few features into Solaris to boost up nfs-client speed: DirectIO and AsyncIO.

As far as I understand Suns techniques DirectIO should bypass filesystem buffering and AsyncIO implements an asynchronious IO while using nfs and not a local filesystem.

Is there a possibility in hpux11.11 to use similar techniques to speed up the nfs-client?

Thanks for your help!
Walter

[harry d brown jr]

I doubt it.

are your nfs clients capable of gig-e speeds? Is your NFS server? If so, go gig-e with jumbo frames.

live free or die
harry d brown jr

[Dave Olker · Accepted solution]

Hi Walter,

NFS version 3, by default, uses asynchronous I/O unless your application opens a file and specifically requests synchronous semantics. In other words, if your application doesn't request synchronous behavior, you get async behavior.

There are certain cases where the client forces synchronous behavior, like when file locking is done, but there are tunable parameters on the client to allow it to continue using asynchronous I/O to locked files - provided the client application locks the entire file. If the application locks only a portion of the file then the client reverts to synchronous I/O.

As for Direct I/O, 11.11 does not offer this. However, 11.23 clients (11i v2) do offer this feature via the "forcedirectio" mount option. If you need Direct I/O with NFS, you'll need to upgrade to 11i v2.

Regards,

Dave

[Walter Kießl]

Hi Dave,

thanks for your answer.

Since I posted this thread I have worked half through your formidable book about optimizing nfs performance on hpux. And thus I have a better understanding of what you have written.

As far as I understand this, Sun's AsyncIO-Implementation should enable the NFS client to do AsyncIO even if the application (in this/my case oracle) opens the file for SyncIO. With this they speed up Oracle database performance without changing the way Oralce opens the database files.

Again thanks for your help and for the great book!

Walter