3.5.1 Matrix Server Administration Guide

Chapter 8: Configure MxFS-Linux 166

If you are receiving I/O errors with a soft mount, you may want to

consider either switching to a hard mount or raising your timeio and/or

retrans parameters to compensate. Consider that the maximum

acceptable time delay for an nfs mount to respond before receiving an I/O

error is (retrans*timeo). In the above example, this is 4*0.7=2.8 seconds.

NFS Reads and Writes

The read and write size that NFS uses to read files from an NFS server or

to write files to an NFS server should be set to 32K. This is done with the

rsize and wsize mount options.

Because of the nature of the NFS protocol, single-stream write

performance is entirely bounded by the size of the I/Os submitted by the

client. Since NFS is “stateless” (disregarding file locking), the only way to

avoid data loss is for writes to actually be committed to storage. By

default, NFS v3 mounts are “synchonous.” If a client is submitting 4K

writes, each write needs to be transmitted, received by the server,

submitted to disk, written, and then a response generated. The latency

causes very low throughput. If the client submits a 1MB write, it will be

broken down into “wsize” writes (32K if the tuning above is performed).

All but the last are immediately acked by the server, and only the final

requires a commit/write, allowing for much higher stream performance.

This can be an issue when using commands such as cp, which use very

small buffer sizes. Use of cp to copy a file to an nfs-mounted filesystem

will yield poor I/O rates. Using dd with a large blocksize to perform the

same operation will give vastly improved results.

DB Optimized Mount Option

The DB Optimized (or DBOPTIMIZE) mount option is separately

licensed and is intended to be used with the MxDB-Oracle HiAv product.

It should not be used for general-purpose NFS access.

sync and async Mount Options

See “Considerations for Ensuring Data Safety” on page 136 for

information about these mount options.