Jiva shows more utilization as can be seen: two possible reasons:
Filesystem metadata upon format (however, this additional usage seems more for a jiva vol compared to a higer sized raw disk attached to the VM. Needs confirmation from Sumit)
Smaller blocks treated as 4k blocks/packing of blocks (however, there is some fs optimization activity here which minimizes this impact for contiguous sequential workload like dd, v/s staggered/random writes)
Might be worth trying out comparing different workloads across local-pv / jiva from same relative start-point (i.e., 20K local PV = 65 MB from jiva and see if same increase in util is observed)
All above questions are important if monitor-pv is used as metrics source for both jiva as well as local PV.
In case of Local PV hostpath & device, monitor-pv mirrors expected app usage via du.
Thoughts based on above observations:
It is better to rely on different metrics sources for different PVs (jiva, cstor, local pv, zfs local pv) -- i.e. grafana panels uses different queries against diff metrics sources.
For jiva/cstor where there is possibility of much divergence b/w user/app usage and actual usage, we can show different graphs or lines to highlight the difference and set the right expectation.
Test Params:
Results:
Notes:
Jiva shows more utilization as can be seen: two possible reasons:
In case of Local PV hostpath & device, monitor-pv mirrors expected app usage via du.
Thoughts based on above observations:
It is better to rely on different metrics sources for different PVs (jiva, cstor, local pv, zfs local pv) -- i.e. grafana panels uses different queries against diff metrics sources.
For jiva/cstor where there is possibility of much divergence b/w user/app usage and actual usage, we can show different graphs or lines to highlight the difference and set the right expectation.