On 10/11/2018 2:19 PM, Wido den Hollander wrote:
On 10/11/2018 12:44 PM, Igor Fedotov wrote:
Wido,
and one more comment on BlueFS allocation issues you're probably
referring to.
IMO this is completely unrelated to allocator implementation - that
rather a design flaw which might cause lack of ***BlueFS*** space at
main device even when the device still has plenty of free space. The
root cause behind this is probably the "lazy" nature of BlueFS rebalance
procedure which runs in background. Under heavy load it might fall behind.
I'm planning to start working on the fix shortly.
Thank you for the feedback, in the meantime I've opened this PR:
https://github.com/ceph/ceph/pull/24542
Suggesting to backport instead:
https://github.com/ceph/ceph/pull/20465
As my thinking was actually that the logging killed the OSD as it died
under the load of the logging.
Wido
Thanks,
Igor
On 10/11/2018 2:52 AM, Wido den Hollander wrote:
Hi,
I was wondering how stable useful the Bitmap allocator is vs the Stupid
allocator with BlueStore under Luminous.
In the master branch the Bitmap one is now the default, but why?
Searching for this I couldn't find a proper reason.
The reason I'm asking is that I'm seeing some issues with BlueStore OSDs
running into BlueFS allocation issues when they are filling up.
(WAL+DB+DATA on same SSD).
So one of the things I wanted to try is to switch to Bitmap, but I'd
like to check the stability and reasoning behind it first.
Thanks,
Wido
