[<prev] [next>] [thread-next>] [day] [month] [year] [list]
Message-ID: <20140103164739.GE7494@localhost>
Date: Sat, 4 Jan 2014 00:47:39 +0800
From: fengguang.wu@...el.com
To: "Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
Cc: LKML <linux-kernel@...r.kernel.org>, lkp@...ux.intel.com
Subject: [rcu] 026ad2835: +30% increased context switches
Hi Paul,
We are running 0day performance tests for both new and old kernels,
and happen to notice that your old commit 026ad2835 ("rcu: Drive
quiescent-state-forcing delay from HZ") increased the context switches
(accidentally?) in the will-it-scale/pread1 test case. Just would like
to confirm with you if this is expected. Attached is the kconfig. The
test machine is a 4S ivybridge system that has 120 CPU threads.
Comparing first bad commit with its parent:
9a5739d73f9369b 026ad2835ce6202069e7aa0b1
--------------- -------------------------
782 ~ 0% +30.1% 1017 ~ 0% vmstat.system.cs
13622 ~ 0% +26.2% 17195 ~ 0% time.involuntary_context_switches
Comparing Linux releases:
v3.10 v3.11 v3.12 v3.13-rc6
--------------- ------------------------- ------------------------- -------------------------
788 ~ 0% +29.9% 1023 ~ 3% +34.9% 1063 ~ 1% +38.6% 1091 ~ 0% vmstat.system.cs
13807 ~ 1% -18.1% 11306 ~40% +24.1% 17128 ~ 0% +25.9% 17376 ~ 0% time.involuntary_context_switches
samples from the bisect good/bad commits
time.involuntary_context_switches
18000 ++-----------------------------------------------------------------+
| O O |
17500 O+ OO O O O O O |
17000 ++O O O O OO O O O O O O |
| |
16500 ++ |
16000 ++ |
| |
15500 ++ |
15000 ++ |
| |
14500 ++ |
14000 *+ |
| *.**. .*.* .*.*.*. .* .*.*.*. *.*. .*.**.*. .*.*.**.*. .* |
13500 ++-----*-*----*-------*--*-------*----*--------*----------*-*--*-*-*
vmstat.system.cs
1050 ++------------------------------------------------------------------+
O O O OO O O O O OO O O O O O O O |
1000 ++ O O O O |
| |
| |
950 ++ |
| |
900 ++ |
| |
850 ++ |
| |
| .*. *. |
800 *+*. .**.*.*.*.*.**.* *.*.**.*. .*. *.*.*.*.*. : *.*. .* .*. |
| * * *.* * *.* * *.*
750 ++------------------------------------------------------------------+
Thanks,
Fengguang
View attachment "x86_64-lkp" of type "text/plain" (80581 bytes)
Powered by blists - more mailing lists