ZFS: logbias

PSARC/2009/423:

Summary

Provide zfs with the ability to control the use of resources used for synchronous (eg fsync, O_DSYNC) requests. In particular it enables substantially better performance for Oracle and potentially other applications.

Background

Oracle manages two major types of files, the Data Files and the Redo Log files. Writes to Redo Log files are in the path of all transactions and low latency is a requirement. It is critical for good performance to give high priority to these write requests.

Data Files are also the subject of writes from DB writers as a form of scrubbing dirty blocks to insure DB buffer availability. Write latency is much less an issue. Of more more importance is achieving an acceptable level of throughput. These writes are less critical to delivered performance.

Both types of writes are synchronous (using O_DSYNC), and thus treated equally. They compete for the same resources: separate intent log, memory, and normal pool IO. The Data File writes impede the potential performance of the critical Redo Log writers.

Proposal

Create a new "logbias" property for zfs datasets.

If logbias is set to 'latency' (the default) then there is no change from the current implementation. If the logbias property is set to 'throughput' then intent log blocks will be allocated from the main pool instead of any separate intent log devices (if present). Also data will be written immediately to spread the write load thus making for quicker subsequent transaction group commits to the pool.

To change the property, an admin will use the standard zfs set command:

# zfs set logbias=latency {dataset}
# zfs set logbias=throughput {dataset}