dm-cache for Fedora 21?

dm-cache for Fedora 21?

[Rolf Fokkens]

Hi all,

For Fedora 20 I packaged bcache-tools so Fedora users are able to use 
bcache. It's not integrated in Anaconda, but users who feel confident on 
the command line will able to use bcache. The next step would be to 
include bcache support in Anaconda, which may happen in Fedora 21.

Fedora 20 seemed  to be too early to include dm-cache support, but I've 
seen a lot of activity on dm-cache on the mail list recently. So maybe 
dm-cache will be "ready" for Fedora 21? AFAIK userland support for 
dm-cache will be included in LVM2, but I don't know what the status is. 
Full integration of dm-cache in Fedora 21 would imply integration in 
Anaconda as well.

And maybe including bot bcache and dm-cache support in Anaconda is a 
little much.

What are your views on the status of dm-cache?

Rolf

Re: dm-cache for Fedora 21?

[Alasdair G Kergon]

We're aiming to have the first drop of lvm code using dm-cache by the
end of this month, and then to gradually fill out and complete the
features from there: expect patches to start to appear on lvm-devel
shortly.

Alasdair

Re: dm-cache for Fedora 21?

[Joe Thornber]

On Sat, Jan 11, 2014 at 03:56:19PM +0100, Rolf Fokkens wrote:
> What are your views on the status of dm-cache?

I know of people outside our development team who've been testing
dm-cache quite thoroughly (eg, soak test runs that go for weeks, 4T
SSD, huge origin etc.).  So my confidence in dm-cache is improving.

There are still some significant things I need to do over the next few
weeks.  Probably the most important being reducing the memory
footprint, and making the cache spot when a working set has changed
and migrate the chunks more quickly.

- Joe

Re: dm-cache for Fedora 21?

[Paul B. Henson]

> Joe Thornber
> Sent: Monday, January 13, 2014 2:18 AM
>
> SSD, huge origin etc.).  So my confidence in dm-cache is improving.

Just out of curiosity, when do you think you will remove the experimental
tag? 3.14? 3.15? Or too soon yet to tell :)?

Re: dm-cache for Fedora 21?

[Joe Thornber]

On Mon, Jan 13, 2014 at 01:01:58PM -0800, Paul B. Henson wrote:
> > Joe Thornber
> > Sent: Monday, January 13, 2014 2:18 AM
> >
> > SSD, huge origin etc.).  So my confidence in dm-cache is improving.
> 
> Just out of curiosity, when do you think you will remove the experimental
> tag? 3.14? 3.15? Or too soon yet to tell :)?

Once I finish the dm-era work (which makes dm-cache play nicely with
snapshots on external devices).  potentially 3.14.

- Joe

Re: dm-cache for Fedora 21?

[Rolf Fokkens]

On 01/14/2014 11:21 AM, Joe Thornber wrote:
> On Mon, Jan 13, 2014 at 01:01:58PM -0800, Paul B. Henson wrote:
>> Just out of curiosity, when do you think you will remove the experimental
>> tag? 3.14? 3.15? Or too soon yet to tell :)?
> Once I finish the dm-era work (which makes dm-cache play nicely with
> snapshots on external devices).  potentially 3.14.
This is all very promising for F21. AFAIK dm-cache can be integrated 
with / applied to existing volumes / device mapper targets which would 
imply that (as opposed to bcache) no migration of existing LVM volumes 
to dm-cache is need. An this would also mean that there's no need to 
integrate it explicitly in Anaconda, since is dm-cache can be applied 
after installation. Is my understanding correct?

Rolf

Re: dm-cache for Fedora 21?

[Joe Thornber]

On Wed, Jan 15, 2014 at 12:24:41AM +0100, Rolf Fokkens wrote:
> On 01/14/2014 11:21 AM, Joe Thornber wrote:
> >On Mon, Jan 13, 2014 at 01:01:58PM -0800, Paul B. Henson wrote:
> >>Just out of curiosity, when do you think you will remove the experimental
> >>tag? 3.14? 3.15? Or too soon yet to tell :)?
> >Once I finish the dm-era work (which makes dm-cache play nicely with
> >snapshots on external devices).  potentially 3.14.
> This is all very promising for F21. AFAIK dm-cache can be integrated
> with / applied to existing volumes / device mapper targets which
> would imply that (as opposed to bcache) no migration of existing LVM
> volumes to dm-cache is need. An this would also mean that there's no
> need to integrate it explicitly in Anaconda, since is dm-cache can
> be applied after installation. Is my understanding correct?

I'm not sure what the LVM interface to caching will look like.  But
I'd expect you to be able to point to any existing volume and say
'speed that up with XG of cache'.

Mike Snitzer tells the 3.14 merge window is opening very soon, so
we'll aim to drop the experimental tag in 3.15.

- joe

Re: dm-cache for Fedora 21?

[Alasdair G Kergon]

On Wed, Jan 15, 2014 at 10:34:46AM +0000, Joe Thornber wrote:
> I'm not sure what the LVM interface to caching will look like.  But
> I'd expect you to be able to point to any existing volume and say
> 'speed that up with XG of cache'.
 
To be clear, I'd qualify that as:
  any existing *Logical Volume*

There are no plans at this stage to support the conversion of existing
volumes created outside LVM.

Alasdair

Re: dm-cache for Fedora 21?

[Paul B. Henson]

> Alasdair G Kergon
> Sent: Wednesday, January 15, 2014 5:22 AM
> 
> On Wed, Jan 15, 2014 at 10:34:46AM +0000, Joe Thornber wrote:
> > I'm not sure what the LVM interface to caching will look like.  But
> > I'd expect you to be able to point to any existing volume and say
> > 'speed that up with XG of cache'.
> 
> To be clear, I'd qualify that as:
>   any existing *Logical Volume*

From a use case perspective, I'd like to be able to point to a pv (or
perhaps a vg), and say 'speed up that pv/vg with XG of cache',  so hopefully
that will be supported by lvm. My tentative layering for testing dm-cache
prior to lvm support was going to be <raw disk><dm-cache><pv>, ie, the pv
was going to be created on the cached device...

Re: dm-cache for Fedora 21?

[Rolf Fokkens]

On 01/16/2014 02:46 AM, Paul B. Henson wrote:
> From a use case perspective, I'd like to be able to point to a pv (or 
> perhaps a vg), and say 'speed up that pv/vg with XG of cache', so 
> hopefully that will be supported by lvm. My tentative layering for 
> testing dm-cache prior to lvm support was going to be <raw 
> disk><dm-cache><pv>, ie, the pv was going to be created on the cached 
> device... -- dm-devel mailing list dm-devel@redhat.com 
> https://www.redhat.com/mailman/listinfo/dm-devel 
Related to my bcache experience I appreciate this way of layering block 
devices. Thought on the other hand I can also appreciate the ability to 
decide per LV to apply caching or not, people did this with bcache as 
well (<raw-disk><pv><vg><lv><bcache>).

For the moment I assume dm-cache is a per LV caching system (since 
nobody confirmed Paul's idea). This leaves me wondering if the cache 
(device) can be shared by multiple LV's. I would really like this, is 
this the case?

Rolf

Re: dm-cache for Fedora 21?

[Paul B. Henson]

> From: Rolf Fokkens
> Sent: Thursday, February 27, 2014 1:49 PM
>
> For the moment I assume dm-cache is a per LV caching system (since
> nobody confirmed Paul's idea). This leaves me wondering if the cache
> (device) can be shared by multiple LV's. I would really like this, is
> this the case?

I followed up on this question on the lvm development list; they pretty much
confirmed that based on the layering and api of lvm, they did not plan to
support using a dm-cache for a pv or vg, only for an lv.

However, one of the developers did give an interesting suggestion combining
the use of dm-cache and thin provisioning, effectively allowing you to use a
single cache for a number of lv's, as long as they are all allocated out of
the same thin pool:

	
https://www.redhat.com/archives/lvm-devel/2014-February/msg00067.html

This looks interesting, while it's not exactly what I was originally
considering, I think it potentially will achieve the same effective goal. I
could create a single thin pool lv the size of the entire backing pv, and
then create thin lv's out of it for my actual production needs, all of which
would be cached by the single ssd pv.

I'm not very familiar with thin provisioning at this time though, so I
really need to do quite a bit of homework on that to make sure there aren't
going to be any issues problematic for my deployment. And, of course, lvm
support for cache devices is still only available in git head and changing
on an almost daily basis :).

I would be interested in any feedback or thoughts anyone might have on this
proposal and its viability and performance, thanks.

Re: dm-cache for Fedora 21?

[Paul B. Henson]

> From: Rolf Fokkens
> Sent: Thursday, February 27, 2014 1:49 PM
>
> For the moment I assume dm-cache is a per LV caching system (since
> nobody confirmed Paul's idea).

Oh, just to clarify, dm-cache itself can be used to cache a pv, sitting
between the raw device and lvm; it's just that lvm will not support setting
that up, you would have to do that directly with dmsetup... I posted an
inquiry on that topic back in December:

https://www.redhat.com/archives/dm-devel/2013-December/msg00038.html

It was pretty much confirmed that it would work, it would just be a rather
manual process.