Hello oVirt Development team

 

Not sure you got a chance to see this email, resending it.

Thanks

Ranjit

From: Ranjit DSouza
Sent: Monday, July 9, 2018 1:16 PM
To: 'Nir Soffer' <nsoffer@redhat.com>
Cc: devel <devel@ovirt.org>; DL-VTAS-ENG-NBU-EverestFalcons <DL-VTAS-ENG-NBU-EverestFalcons@veritas.com>; Navin Tah <Navin.Tah@veritas.com>; Sudhakar Paulzagade <Sudhakar.Paulzagade@veritas.com>; Pavan Chavva; Yaniv Lavi (Dary) <ylavi@redhat.com>; Nisan, Tal <tnisan@redhat.com>; Daniel Erez <derez@redhat.com>
Subject: RE: [EXTERNAL] Re: [ovirt-devel] Image Transfer mechanism queries/API support

 

Hi Nir

 

Thanks for getting back!

We have few follow up questions:

 

1.       On the new ‘Allocated extents API’:

Can you share the release timeline for 4.2.z?  From the link below it seems like it will be available on 7/30/2018.  

https://www.ovirt.org/develop/release-management/releases/4.2.z/release-management/

However, we thought we would double check on this.

2.       If this will be available in 4.2.z, does It mean, we can assume it will be backported to 4.3 also?

3.       When we downloaded the snapshot disk using Image Transfer API, the resulting format of the disk is “raw”.  However, for upload, we must upload a qcow2 disk (to enable further snapshots).

It means, we need to convert it first using qemu-img convert. Or is there a way we directly ask via API for a qcow2 instead?

 

Portion of the response of “GET /storagedomains/{storagedomain:id}/disksnapshots”

    "format" : "raw",

    "shareable" : "false",

    "sparse" : "true",

    "status" : "ok",

    "snapshot" : {

      "id" : "4756036e-92aa-4ebb-ae4b-052a30cd5109"

    },

    "actual_size" : "1345228800",

    "content_type" : "data",

    "propagate_errors" : "false",

    "provisioned_size" : "21474836480",

    "storage_type" : "image",

    "total_size" : "0",

    "wipe_after_delete" : "false",

 

4.       When downloading a snapshot in chunks, Is there any recommended chunk size? For our study we used 512 KB. Checked the documentation too.

5.       Regarding your ask on ‘Can you file RFE for this, explaining the use case and the current performance’.

Since you do not recommend direct access to NFS storage, we will consider it only if we see significant performance degradation using http download.  

Also, if time permits, we may check if there is a significant performance benefit using Unix socket, over http download (via Rest API).

But as it stands now, getting the allocated extents support (soon), will alleviate most of our performance concerns.

 

Thanks

Ranjit

 

From: Nir Soffer [mailto:nsoffer@redhat.com]
Sent: Tuesday, July 3, 2018 4:29 PM
To: Ranjit DSouza <Ranjit.DSouza@veritas.com>
Cc: devel <devel@ovirt.org>; DL-VTAS-ENG-NBU-EverestFalcons <DL-VTAS-ENG-NBU-EverestFalcons@veritas.com>; Navin Tah <Navin.Tah@veritas.com>; Sudhakar Paulzagade <Sudhakar.Paulzagade@veritas.com>; Pavan Chavva; Yaniv Lavi (Dary) <ylavi@redhat.com>; Nisan, Tal <tnisan@redhat.com>; Daniel Erez <derez@redhat.com>
Subject: [EXTERNAL] Re: [ovirt-devel] Image Transfer mechanism queries/API support

 

On Tue, Jul 3, 2018 at 11:32 AM Ranjit DSouza <Ranjit.DSouza@veritas.com> wrote:

...

We had a conversation with Pavan Chavva for supporting RHV. He had suggested to contact you with queries related to oVirt APIs we plan to use.

We have following queries:

 

1.       While downloading a snapshot disk, can we identify allocated extents and download only those using oVirt API? We are able to download the disk using the Image Transfer API mechanism.

However, this method downloads the entire disk including the non-allocated extents, which is a performance overhead. If this functionality does not exist at this point will it be available in near future?

 

Hi Ranjit,

 

There is no way to do this in current 4.2, but we plan to introduce in in 4.2.z.

 

The API will be something like:

 

GET /images/xxx-yyy/map

...

[{ "start": 0, "length": 65536, "depth": 0, "zero": false, "data": true, "offset": 0},

{ "start": 65536, "length": 983040, "depth": 0, "zero": true, "data": false, "offset": 65536},

{ "start": 1048576, "length": 65536, "depth": 0, "zero": false, "data": true, "offset": 1048576},

{ "start": 1114112, "length": 983040, "depth": 0, "zero": true, "data": false, "offset": 1114112},

...

{ "start": 5465571328, "length": 22675456, "depth": 0, "zero": false, "data": true, "offset": 5465571328},

{ "start": 5488246784, "length": 954138624, "depth": 0, "zero": true, "data": false, "offset": 5488246784},

{ "start": 6442385408, "length": 65536, "depth": 0, "zero": false, "data": true, "offset": 6442385408}]

 

This is basically what you get using qemu-img map.

 

You can test play this with this using:

 

    virt-builder Fedora-27 -o /var/tmp/fedora-27.img

    qemu-img map -f raw --output json /var/tmp/fedora-27.img

 

This is the first data segment:

{ "start": 0, "length": 65536, "depth": 0, "zero": false, "data": true, "offset": 0}

 

This is a hole between the first data segment and the second:

{ "start": 65536, "length": 983040, "depth": 0, "zero": true, "data": false, "offset": 65536}

 

This is the second data segment:

{ "start": 1048576, "length": 65536, "depth": 0, "zero": false, "data": true, "offset": 1048576}

 

Based on this output, you will be able to get the allocated parts of the image using:

 

Request:

 

    GET /image/xxx-yyy HTTP/1.1

    Range: bytes=0-65535

 

Response:

 

    HTTP/1.1 206 Partial Content

    Content-Range: bytes 0-65535/6442450944

    

    <data of first segment>

 

Request:

 

    GET /image/xxx-yyy HTTP/1.1

    Range: bytes=1048576-1114111

 

Response:

 

    HTTP/1.1 206 Partial Content

    Content-Range: bytes 1048576-1114111/6442450944

 

    <data of second segment>

 

And so on.

 

If you create a sparse file on your backup media, and download and write

the data segments at the correct offset, you will get the a sparse version of

the image as on the server side.

 

This will work for raw or qcow2 images on NFS >= 4.2, or for qcow2 images on block

storage.

 

For older NFS versions, or raw images on block storage, we can solve the issue by

reading the entire image and detecting zeroes - which is quite expensive, so I'm not

sure we will implement this, maybe it will be done later.

 

We have experimental patch using special sparse format, that can support this use 

case, downloading entire image in one pass. This commit message explain the 

format:

https://gerrit.ovirt.org/#/c/85413/12//COMMIT_MSG

 

For more info on using random I/O APIs, see:

http://ovirt.github.io/ovirt-imageio/random-io

(available since 4.2.3)

 

For example code uploading sparse images see:

https://github.com/oVirt/ovirt-imageio/blob/master/examples/upload

 

For best performance, you should run your application on a oVirt host, using unix

socket to communicate with imageio. See:

http://ovirt.github.io/ovirt-imageio/unix-socket

(will be available in 4.2.5)

 

All this will work only for the non-active layer in a qcow2 chain. We are working now

on incremental backup which will allow the same for the active layer with a running

vm. This is expected in 4.3, and we may have a tech preview at some point in 4.2.z.

Incremantal backup will use the similar API, allowing detection of dirty parts of an image,

so you can download only the data that was changed since the last backup.

Please watch and comment on the feature page:

https://ovirt.org/develop/release-management/features/storage/incremental-backup/

 

We are also considering exposing images using NBD. This will allow downloading

and uploading images using qemu-img from any host. This work depends on TLS-PSK

support in qemu-img and qemu-nbd. You can follow this work here:

https://lists.nongnu.org/archive/html/qemu-devel/2018-06/threads.html#08491

 

2.       Is there an alternate method to transfer a snapshot to and from RHV storage? Are there other methods such as NFS share where we can download snapshot image to and from RHV storage?

We don't support direct access to storage by 3rd party. You should use imageio API.

 

Can you file RFE for this, explaning the use case and the current performance 

issues you experience?

 

Nir