[LU-16875] spinlock contention at pagecache during buffered ior test Created: 05/Jun/23  Updated: 05/Jun/23

Status: Open
Project: Lustre
Component/s: None
Affects Version/s: Lustre 2.15.2
Fix Version/s: None

Type: Bug Priority: Minor
Reporter: Sergey Kachkin Assignee: WC Triage
Resolution: Unresolved Votes: 0
Labels: None
Environment:

Rocky linux 8.7-ib, Lustre 15.2.2, 1 OSS / 4 OST , single client


Attachments: PNG File image-2023-06-06-00-23-09-738.png     File ior_buffered.svg     File ior_odirect.svg    
Epic/Theme: Performance
Severity: 3
Rank (Obsolete): 9223372036854775807

 Description   

Lustre is configured with small configuration 1 OSS /4 OST 

During single threaded buffered test with IOR I noticed high system CPU usage and spinlock contention at between IOR and ptlrpcd. 

buffered ior executed like below: 

ior  -F -w -b 64g -t 256m -e -o /lustre/

 

This seems expected that during the buffered  test there is high pressure on pagecache but I wonder if this spinlock contention between IOR and ptlrpcd can be somehow avoided. 

Additionally I did direct IO test and as expected I do not see any lock contention but 

direct io test: 

ior -b 64g -t 256m -e -w -r -o /lustre/iorfile --posix.odirect

 

During direct IO test there is no contention (as expected) but IOR still consumes a lot of CPU time due to memory operations.  


Generated at Sat Feb 10 03:30:44 UTC 2024 using Jira 9.4.14#940014-sha1:734e6822bbf0d45eff9af51f82432957f73aa32c.