LU-17251 osp: start OST object precreate earlier
If the OST object precreate count gets large (usually due to high
MDT file create workload, but sometimes also forced during testing)
then send an OST_CREATE RPC sooner when the number of precreated
objects gets low.
Currently the MDS will wait until 1/2 of the precreated OST objects
are consumed, but if create_count = 10000, then this can put bursty
create workloads on the OST. Instead, send an OST_CREATE RPC when
the precreate pool is at most 1024 objects below target, so that the
MDS keeps its precreated pool more full and the OST doesn't have to
create so many objects at once (which also locks object directories
for a longer time).
Don't set opd_force_creation=true when osp.*.create_count is set
larger, and instead rely on the improved precreate check to force
OST object creation to start sooner, as opd_force_creation=true
can cause the OSP precreation to stop completely in some cases.
Test-Parameters: testlist=sanity env=ONLY=1-130,HONOR_EXCEPT=y
Test-Parameters: testlist=sanity env=ONLY=1-130,HONOR_EXCEPT=y
Test-Parameters: testlist=sanity env=ONLY=1-130,HONOR_EXCEPT=y
Test-Parameters: testlist=sanity env=ONLY=1-130,HONOR_EXCEPT=y
Test-Parameters: testlist=parallel-scale env=ONLY=rr_alloc,ONLY_REPEAT=10
Test-Parameters: testlist=parallel-scale env=ONLY=rr_alloc,ONLY_REPEAT=10
Test-Parameters: testlist=parallel-scale env=ONLY=rr_alloc,ONLY_REPEAT=10
Test-Parameters: testlist=parallel-scale env=ONLY=rr_alloc,ONLY_REPEAT=10
Fixes:
df5b4c0a8b ("LU-17251 osp: force precreate if create_count grows")
Signed-off-by: Andreas Dilger <adilger@whamcloud.com>
Change-Id: Id2d12636d535485919ca5eec3adb18b1e6ce7057
Reviewed-on: https://review.whamcloud.com/c/fs/lustre-release/+/53245
Tested-by: jenkins <devops@whamcloud.com>
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: Alex Deiter <alex.deiter@gmail.com>
Reviewed-by: Alex Zhuravlev <bzzz@whamcloud.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>