ZFS read performance anomaly / performance regression

[Garrett Wollman]

<<On Sat, 25 Jan 2014 11:43:28 -0800, Matthew Ahrens <mahrens at delphix.com> said:

> How are you reading the files?  Could you be reading in less than
> block-size chunks (default 128K)?

Nope.  dd bs=1024k.

> This sounds similar to issue discussed in the thread with subject "ZFS slow
> reads for unallocated blocks", which was comparing sparse vs cached reads.
>  Since your storage is so fast, prefetch will probably read and cache the
> blocks by the time you issue the read() syscall, so you may be hitting a
> similar issue.

How would that have changed between 9.1 and 9.2?

-GAWollman