The parameter ”order“ doesn't work properly in RBD backend

[luqitao]

There are two parameters can affect the performance of image upload, one is rbd_store_chunk_size, one is order

At present, the value of order by the following formula: order = int (math.log (rbd_store_chunk_size 1024 * 2, 2))
The block size rbd_store_chunk_size is glance every time you read the image, order CEPH is stored in the mirror when the size of the block

For example:
rbd_store_chunk_size=8 order=23 glance 8M, then each read data to the back-end, CEPH is also based on the 8M data of each disk write
rbd_store_chunk_size=32 order=25 glance 32M, then each read data to the back-end, CEPH is also based on the 32M data of each disk write

This way can not play the performance of CEPH, if "order" adjust to a fixed value 23:
rbd_store_chunk_size=32 order=23 glance 32M, then each read data to the back-end CEPH in each 8M, 4 times of concurrent write disk, the actual test that can speed up process of image uploading.

[Kairat Kushaev (kkushaev)]

Could you please specify how it can be fixed? Is it specified in ceph configuration?

[Kairat Kushaev (kkushaev)]

We are calculating order here:
https://github.com/openstack/glance_store/blob/master/glance_store/_drivers/rbd.py#L380
So looks like we are trying to specify it in accordance with rbd_store_chunk_size.
It is not clear how can it be fixed and how we can receive this value from rbd.

[luqitao (luqt2011)]

We can set the order parameter to a fixed value of 23 (or in the configuration file: /etc/cinder/cinder.conf), no need to calculating it, so that it can improve the CEPH's concurrent processing capability.