On Thu, Mar 29, 2018 at 5:19 PM Nir Soffer <nirsof@gmail.com> wrote:
On Thu, Mar 29, 2018 at 1:59 PM Richard W.M. Jones <rjones@redhat.com> wrote:
I found another problem which is sort of related to this thread.  I
hit a time-out waiting for the disk to be unlocked after creation:

https://github.com/libguestfs/libguestfs/commit/8081f54105bd990233f166170890192c1fd7d1f3#diff-5ca47c29ae13efa3959b8b28cf4dbd31R112
line 112

The disk size in this case is 40 GB and we currently wait 5 minutes.

On the ovirt-node there's an ‘fallocate’ process (still) running.
It's taken 23 minutes so far.

This code is using posix_fallocate and it can be very slow on NFS < 4.2,
writing one byte to every 4k block.

Actually it is worse - today we:

1. create virtual_size bytes volume
2. preallocate entire image using posix_fallocate, possibly writing 
   one byte to every 4k sector
3. write virtual_size bytes of data to volume

The preallocation is mostly useless. It can save you time if there
is not enough storage to copy the volume, but you have to pay for
something that practically never happen for every volume you create.

I have this patch:
https://gerrit.ovirt.org/#/c/86282/

With this you will be able to create a raw preallocated
volume without preallocating anything using initial_size=0,
avoiding the slow creation you experience now.

The file will be allocated by uploading the data into the volume.

This will require also changes on engine side, currently we don't 
send initial_size to vdsm for file storage.

It would be nice if you file a performance bug for this issue.

Nir
 
 
Because we didn't reach a satisfactory conclusion of the whole
sparse/format business, I am currently setting sparse = False
(format is always raw in my testing):

https://github.com/libguestfs/libguestfs/commit/8081f54105bd990233f166170890192c1fd7d1f3#diff-5ca47c29ae13efa3959b8b28cf4dbd31R92
line 92

I guess this is the cause of this problem in some sense.  Is there
something we can do?  I'm still unclear how to fix the sparse/format
mess, it seems unsolvable from the client side because we don't have
sufficient information to make a decision.

I don't think always using sparse=False is a good solution. We have an issue
only with block storage, so there is no need to force sparse=False for every
kind of storage.

I think the behavior should be:

- if the user wants raw image (sparse=False) on file domain we can warn that 
creating preallocated disk on file storage can take lot of time, and we should
wait until the disk is ready. Or we can always use thin for upload to file based
storage, this is what we for uploading from the UI.

- if the user wants thin image on block storage, we must allocate the entire
image when we create a disk. In 4.2.4 we will support engine API to reduce
an image to the optimal size, and you will be able to call this when you complete
the upload.

- if the user wants thin image on file storage, or raw image on block storage,
we don't have any issue

In 4.3 we plan to improve engine extend volume api so it will update the ticket
size when extending a disk. With this you will be able to create a thin disk on
block storage and extend it when needed.

Nir