From mboxrd@z Thu Jan 1 00:00:00 1970 MIME-Version: 1.0 References: <909d4cae-ddd2-3951-eee8-8dec8faa6f22@redhat.com> In-Reply-To: From: =?UTF-8?Q?Tomas_Dalebj=C3=B6rk?= Date: Wed, 23 Oct 2019 13:24:31 +0200 Message-ID: Content-Type: multipart/alternative; boundary="000000000000c6a6b505959229f0" Subject: Re: [linux-lvm] exposing snapshot block device Reply-To: LVM general discussion and development List-Id: LVM general discussion and development List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , List-Id: To: Gionatan Danti Cc: Zdenek Kabelac , LVM general discussion and development --000000000000c6a6b505959229f0 Content-Type: text/plain; charset="UTF-8" I have tested FusionIO together with old thick snapshots. I created the thick snapshot on a separate old traditional SATA drive, just to check if that could be used as a snapshot target for high performance disks; like a Fusion IO card. For those who doesn't know about FusionIO; they can deal with 150-250,000 IOPS. And to be honest, I couldn't bottle neck the SATA disk I used as a thick snapshot target. The reason for why is simple: - thick snapshots uses sequential write techniques If I would have been using thin snapshots, than the writes would most likely be more randomized on disk, which would have required more spindles to coop with this. Anyhow; I am still eager to hear how to use an external device to import snapshots. And when I say "import"; I am not talking about copyback, more to use to read data from. Regards Tomas Den ons 23 okt. 2019 kl 13:08 skrev Gionatan Danti : > On 23/10/19 12:46, Zdenek Kabelac wrote: > > Just few 'comments' - it's not really comparable - the efficiency of > > thin-pool metadata outperforms old snapshot in BIG way (there is no > > point to talk about snapshots that takes just couple of MiB) > > Yes, this matches my experience. > > > There is also BIG difference about the usage of old snapshot origin and > > snapshot. > > > > COW of old snapshot effectively cuts performance 1/2 if you write to > > origin. > > If used without non-volatile RAID controller, 1/2 is generous - I > measured performance as low as 1/5 (with fat snapshot). > > Talking about thin snapshot, an obvious performance optimization which > seems to not be implemented is to skip reading source data when > overwriting in larger-than-chunksize blocks. > > For example, consider a completely filled 64k chunk thin volume (with > thinpool having ample free space). Snapshotting it and writing a 4k > block on origin will obviously cause a read of the original 64k chunk, > an in-memory change of the 4k block and a write of the entire modified > 64k block to a new location. But writing, say, a 1 MB block should *not* > cause the same read on source: after all, the read data will be > immediately discarded, overwritten by the changed 1 MB block. > > However, my testing shows that source chunks are always read, even when > completely overwritten. > > Am I missing something? > > -- > Danti Gionatan > Supporto Tecnico > Assyoma S.r.l. - www.assyoma.it > email: g.danti@assyoma.it - info@assyoma.it > GPG public key ID: FF5F32A8 > --000000000000c6a6b505959229f0 Content-Type: text/html; charset="UTF-8" Content-Transfer-Encoding: quoted-printable
I have tested FusionIO together with old thick snapsh= ots.
I created the thick snapshot on a separate old traditional S= ATA drive, just to check if that could be used as a snapshot target for hig= h performance disks; like a Fusion IO card.
For those who doesn&#= 39;t know about FusionIO; they can deal with 150-250,000 IOPS.

And to be honest, I couldn't bottle neck the SATA disk= I used as a thick snapshot target.
The reason for why is simple:=
- thick snapshots uses sequential write techniques
If I would have been using thin snapshots, than the writes woul= d most likely be more randomized on disk, which would have required more sp= indles to coop with this.

Anyhow;
I am s= till eager to hear how to use an external device to import snapshots.
=
And when I say "import"; I am not talking about copyback, mo= re to use to read data from.

Regards Tomas

De= n ons 23 okt. 2019 kl 13:08 skrev Gionatan Danti <g.danti@assyoma.it>:
On 23= /10/19 12:46, Zdenek Kabelac wrote:
> Just few 'comments' - it's not really comparable - the eff= iciency of
> thin-pool metadata outperforms old snapshot in BIG way (there is no > point to talk about snapshots that takes just couple of MiB)

Yes, this matches my experience.

> There is also BIG difference about the usage of old snapshot origin an= d
> snapshot.
>
> COW of old snapshot effectively cuts performance 1/2 if you write to <= br> > origin.

If used without non-volatile RAID controller, 1/2 is generous - I
measured performance as low as 1/5 (with fat snapshot).

Talking about thin snapshot, an obvious performance optimization which
seems to not be implemented is to skip reading source data when
overwriting in larger-than-chunksize blocks.

For example, consider a completely filled 64k chunk thin volume (with
thinpool having ample free space). Snapshotting it and writing a 4k
block on origin will obviously cause a read of the original 64k chunk,
an in-memory change of the 4k block and a write of the entire modified
64k block to a new location. But writing, say, a 1 MB block should *not* cause the same read on source: after all, the read data will be
immediately discarded, overwritten by the changed 1 MB block.

However, my testing shows that source chunks are always read, even when completely overwritten.

Am I missing something?

--
Danti Gionatan
Supporto Tecnico
Assyoma S.r.l. - www.assyoma.it
email: g.danti@assy= oma.it - info@assy= oma.it
GPG public key ID: FF5F32A8
--000000000000c6a6b505959229f0--