[zfs-discuss] Re: FlashCache and ZFS

Greg landgre at gmail.com
Tue Jun 26 20:29:09 EDT 2012


For something FS agnostic (or OS or kernel agnostic, for that matter)
look at something like LSI CacheCade (http://thessdreview.com/raid-
enterprise/lsi-megaraid-cachecade-pro-2-0-review-total-storage-
acceleration-realized/) or Fusion-io directCache (http://
www.fusionio.com/data-sheets/directcache/).  You do have spring for
both hw & sw licensing, but worth consideration for enterprise
applications.

On Jun 25, 5:37 am, Aneurin Price <aneurin.pr... at gmail.com> wrote:
> On 25 June 2012 10:30, Ryan How <r... at exemail.com.au> wrote:
>
> > Isn't it a write-back cache to improve write performance?
> > Like bcache (http://bcache.evilpiepirate.org/)
>
> > Basically creating a bigger write buffer by using an ssd.
>
> Both of those appear to need a block device to use directly (and in
> the case of bcache, a block device initialised before starting to use
> it) so they'd only be useful if you were using zvols with another FS
> on top (eg. zvol->bcache->ext4).
>
> That might be worthwhile, to be sure, but even better would be
> something that could be dropped into an existing ZFS system - and that
> would need to be specifically written for ZFS.
>
> Actually, I can sort of envision something completely FS-agnostic that
> would do the job, by using a unioned filesystem that writes first to a
> fast device, and then moves the data asynchronously to the underlying
> device - using some kind of unionfs would mean it could be done at a
> higher level so it wouldn't need access to the underlying block
> device, but it would have different failure and performance
> characteristics of course. Plus there's the drawback that, so far as I
> know, it doesn't exist :P. Or maybe it does - anyone know?
>
> This is something I've wished for quite a lot - I have some datasets
> with dedup enabled which are excruciatingly slow to write to once the
> in-memory buffers are filled (ie. after about ten seconds). I'd love
> to be able to say 'take this data until this large disk buffer is
> full; I don't care if you then need to spend an hour writing it back
> and deduplicating it'.



More information about the zfs-discuss mailing list