Re:Super Fast Local Workloads with LXD, ZFS, and Juju
by justinsaccount on Friday February 12, 2016 @11:16AM
(#11088258)
Yeah.. it is a subtle difference between
and
[ Parent
| on HN ]
Re:Super Fast Local Workloads with LXD, ZFS, and Juju
by dr_zoidberg on Friday February 12, 2016 @11:34AM
(#11088392)
Could you elaborate on the semantics of "cache"? I do understand the mirror keyword, but I know next to nothing about ZFS and don't really get what "cache" means, considering bouth your and OPs comment.
Re:Super Fast Local Workloads with LXD, ZFS, and Juju
by clord on Friday February 12, 2016 @12:00PM
(#11088599)
a cache drive is l2arc. Brendan Gregg's post[0] explains it well, but the TLDR is that l2arc is very much like an l2 cache in a cpu. Recent and relevant items are held on a fast drive in case it will be needed in the near future. It is really an L2 cache: L1 is called ARC and is held in dram.
If I recall correctly, the extra indexes and tags mean l2arc requires more memory to work efficiently.
Re:Super Fast Local Workloads with LXD, ZFS, and Juju
by semi-extrinsic on Friday February 12, 2016 @03:44PM
(#11090399)
When your "medium-sized test server" has 128 GB RAM and half a TB worth of SSDs... in 2008! Damn.
Re:Super Fast Local Workloads with LXD, ZFS, and Juju
by justinsaccount on Friday February 12, 2016 @11:54AM
(#11088548)
You can use a fast disk like an SSD as a 2nd level cache device after ram. So reads will first check the main ram cache first, then the ssd cache disk before finally hitting the underlying storage.
Re:Super Fast Local Workloads with LXD, ZFS, and Juju
by feld on Friday February 12, 2016 @12:12PM
(#11088678)
L2ARC is extremely well documented right in the code
https://svnweb.freebsd.org/base/head/sys/cddl/contrib/openso...
Re:Super Fast Local Workloads with LXD, ZFS, and Juju
by abrookewood on Friday February 12, 2016 @05:08PM
(#11091046)
Correct me if Im wrong, but he just built a pool with 3 layers of parity right? The first drive with three drives mirroring it? Where as the later is effectively RAID10, a stripe across 2 mirrored vdevs.