Achieving the Best NAS Performance Possible

Now that you know a few of the considerations you'll need to make before moving your legacy applications on to a cluster filesystem that uses the NFS protocol, note a few of the things you can do to make your cluster filesystem perform optimally:

• Dedicate a network to isolate NFS traffic.

• Use a high-quality NAS server from a NAS vendor (these NAS servers use nonvolatile RAM to commit NFS write operations as quickly as possible).

• Play with the read and write sizes used on the NFS client (see "wsize" and "rsize" on the mount man page). When using NFS over TCP (discussed in a moment), the recommended read and write size is 32K. (The read and write size should always be larger than the NFS client's page size, which is usually 4K.) Sending larger read and write requests from an NFS client can significantly reduce the latency of NFS operations; however, you can't always control the size of NFS packets by simply changing these numbers (the application program that uses NFS may also need to be modified).

• Use specialized networking techniques from your NAS vendor (such as trunking multiple network connections) to remove network performance bottlenecks.^[23]

Note 

In an effort to boost the NFS server's performance, the current default Linux NFS server configuration uses asynchronous NFS. This lets a Linux NFS server (acting as an NAS device) commit an NFS write operation without actually placing the data on the disk drive. This means if the Linux NFS server crashes while under heavy load, severe data corruption is possible. For the sake of data integrity, then, you should not use async NFS in production. An inexpensive Linux box using asynchronous NFS will, however, give you some feel for how well an NAS system performs.^[24]

Don't assume you need the added expense of a GigE network for all of your cluster nodes until you have tested your application using a high-quality NAS server.

In Chapter 18, we'll return to the topic of NFS performance when we discuss how to use the Ganglia monitoring package.

^[23]Although, as we've already discussed the network is not likely to become a bottleneck for most applications

^[24]Though async NFS on a Linux server may even outperform a NAS device.