Bug #7521
closedAdd more events (hold object context) to OpTracker to better analyze performance
0%
Description
This ticket is to help on the effort I sent to community for this - http://www.spinics.net/lists/ceph-users/msg07866.html
When I was analyzing read performance, it turned out a large number of read / getattrs ops took quite a long time before 'reached_pg' and 'started' (via history ops), and further analysis showed the time were spent building the object context, which did filestore i/o the first time (getattr "_"), as the first time I/O would do FileStore::lfn_open the first time in the transaction, it is likely to miss the fd cache thus need to file lookup and open, both steps turned out spending long time for cold data.
It would be nice to add more events to assist for such analysis, e.g. mark the event getting the object context between 'reached_pg' and 'started'.
Are people interested in such fine grain events? I can create a pull request for such.
Updated by Guang Yang over 9 years ago
- Status changed from New to Won't Fix
With more understanding of the tracker, I found actually the issue being tracked by this but can be achieved by the current tracker. Close a won't fix.