[SERVER-12096] Enterprise rhel62 build is much slower than rhel57 build in GSSAPI Connection Performance Created: 14/Dec/13  Updated: 19/May/14  Resolved: 17/Jan/14

Status: Closed
Project: Core Server
Component/s: Security
Affects Version/s: 2.4.8, 2.5.3
Fix Version/s: None

Type: Bug Priority: Major - P3
Reporter: Rui Zhang (Inactive) Assignee: Rui Zhang (Inactive)
Resolution: Won't Fix Votes: 0
Labels: 26qa
Remaining Estimate: Not Specified
Time Spent: Not Specified
Original Estimate: Not Specified

Issue Links:
Related
Operating System: ALL
Participants:

 Description   

During my authentication perf test, found that rhel6.2 build is much slower in setting up GSSAPI connection than rhel5.7 build, and take much more CPU.

This behavior is the same for 2.4.8 as well. In order to make sure this is not due to environment, I setup two setups, one in EC2 and one in my local environment.

Here is output from EC2 (both server run with m1.xlarge)

2.5.5-pre (2013-12-11) rhel62 vs rhel57:
2.5.5-pre GSSAPI rh62/ec2 avg_time uses 332.926730102% more than 2.5.5-pre GSSAPI rh57/ec2 | 24.2043759705 vs 5.59087122313
2.5.5-pre GSSAPI rh62/ec2 CPU uses 42.4165090943 % more than 2.5.5-pre GSSAPI rh57/ec2 | 94.9427373159 vs 66.6655417407

the same behavior also presents with 2.4.8
2.4.8 rhel62 vs rhel57:
2.4.8 GSSAPI rh62/ec2 avg_time uses 324.64563971 % more than 2.4.8 GSSAPI rh57/ec2 | 24.6068780705 vs 5.79468520797
2.4.8 GSSAPI rh62/ec2 CPU uses 47.3761253618 % more than 2.4.8 GSSAPI rh57/ec2 | 94.6278591549 vs 64.2084048027

similar from run from local VMs
2.5.5-pre GSSAPI avg_time uses 137.642469807 % more than 2.5.5-pre GSSAPI rh57 | 9.49420886465 vs 3.99516503609
2.5.5-pre GSSAPI CPU uses 134.529540316 % more than 2.5.5-pre GSSAPI rh57 | 68.138736952 vs 29.0533707865



 Comments   
Comment by Rui Zhang (Inactive) [ 19/Dec/13 ]

schwerin will do, I am trying to wrap up Windows platform testing, after it is done, will collect more details for this ticket and SERVER-12087

Comment by Andy Schwerin [ 19/Dec/13 ]

rui.zhang, can you and Davide use the gperftools CPU profiler or the valgrind callgrind tool to collect profiles of the benchmark running on the two platforms? I suspect this issue is beyond our control, but a profile graph would help.

Comment by Rui Zhang (Inactive) [ 19/Dec/13 ]

Re-assign this to schwerin, please re-assign if this shall go to another engineer. Thanks

Comment by Daniel Pasette (Inactive) [ 19/Dec/13 ]

What is the test doing?

Generated at Thu Feb 08 03:27:36 UTC 2024 using Jira 9.7.1#970001-sha1:2222b88b221c4928ef0de3161136cc90c8356a66.