Description of problem:
when I do a simultaneous lookups using ls from multiple clients, the time taken to complete the listing by the last client takes about 11 seconds(other clients varying from 0.5sec to 8sec). this is in case of a disperse volume mounted on 4 fuse clients.
However If i disable other-eager-lock, the response time is almost similar and immediate for all 4 clients ie less than 1 second
This is a serious drop in performance given that the root of volume had about only 4 directories to be looked up for.
note that this discussion has been going on for sometime in BZ#1530519
Version-Release number of selected component (if applicable):
Steps to Reproduce:
1.create an ec volume say 2x(4+2)
2.mount volume on a client and create 4 entries say 4 directories on root
3.now mount volume on another 3 clients
4. from all clients parallelly issue ls
5. now turn of other-eager-lock and re-issue ls parallelly
while with step 5 the response by all clients is similar and less than a second,
with step4 the response time for the last client is as huge as 11sec
Turn off other-eager lock by default
Created attachment 1438519 [details]
screenshot with showing the performance with 3.12.2-11 build
The plan is to turn of eager lock off by default using 1611151 and only turn it on only in cases where it is needed.
To debug the real issue with other.eager-lock enable, I would again request QE to test it with latest release as there are some patches related to performance and that could have fixed the issue.
There was nothing in EC which would have impacted the performance. As per comment #25 we are seeing high inodelk latency on brick side.
1. Do we have plan to enable other eager lock in future?
2. If we do enable it for certain cases, then do we a perf hit? If not keeping this bug open doesn't make much sense to me.