JinHak Kim from NetApp recently reported a similar issue, and shared the results in NetApp internal mailing list.
You might be talking about the same issue, but it's worth sharing his conclusions:
> To make long story short, through the internal test and communication, I concluded that copy-offload works both NFSv3 & NFSv4 protocol.
> When creating a VM, after downloading the glance image to cinder backend via copy-offload, the operation of converting the image format occurs.
> At this time, network IO happens. Don't misunderstand this for glance image regular download.
So the answers are the following:
- Test#1: Creating nova with an image on glance_vol to cinder_vol#1 using copy offload failed. The Image was copied using regular downloading and registered in cache in cinder_vol#1.
Please check if copy offload really failed. In the cinder volume log, if the copy offload tool is returning 0 for the volume you're creating, copy offload is happening correctly in the 1st attempt as well.
It's probably taking longer because in the 1st attempt it requires image conversion.