Merge Spectra Logic's optimizations to the DMU

[Rudd-O]

http://www.youtube.com/watch?v=LtY3vpX-cdM

[Ukko-Ylijumala]

Published on May 19, 2012

Optimizing ZFS for Block Storage

The ZFS file system has been heavily tuned for workloads where file rewrite activity is minimal or is aligned and sized to match ZFS's native record size. Exporting ZFS storage to block consumers, however, presents a situation where every write is rewriting an existing block, and unaligned writes incur a performance killing synchronous read.

This paper and talk presents Spectra Logic's optimizations to ZFS's data management layer (DMU) to convert the majority of these synchronous reads to be asynchronous and, for sequential access patterns, to avoid them entirely. We also describe a new scheme that allows concurrent reads to be issued through the DMU without the need to allocate a thread context for each I/O. The result, as implemented and tested using the FreeBSD operating system, is up to a five fold performance increase for unaligned write workloads and a three fold improvement for random read workloads.

Interesting. This has relevance in the context of bug #361, and maybe also #824.

However, this is already covered by #1074.

[behlendorf]

This code has been around for some time and it's absolutely a good idea. However, Spectra Logic's optimizations haven't yet made it in to FreeBSD proper or Illumos. Someone needs to spend the time to gather up their changes and post them in a form they can be reviewed and tested.

[Rudd-O]

OK, so they haven't been merged. I'll retitle this issue.

[behlendorf]

Closing, these changes were never push upstream and aren't available for us to include.