Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1631847 > unrolled thread

[GIT PULL rcu/urgent] Auto-expedite initial SRCU grace periods

Back to article view | Back to linux.kernel

Contents

  [GIT PULL rcu/urgent] Auto-expedite initial SRCU grace periods "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> - 2017-04-27 05:50 +0200
    Re: [GIT PULL rcu/urgent] Auto-expedite initial SRCU grace periods Ingo Molnar <mingo@kernel.org> - 2017-04-27 08:20 +0200

#1631847 — [GIT PULL rcu/urgent] Auto-expedite initial SRCU grace periods

Hello, Ingo,

This series greatly reduces the performance degradation of Tree SRCU
on a CPU-hotplug stress test.  The effect was not subtle:  Mike Galbraith
measured Classic SRCU at 55 seconds and Tree SRCU at more than 16 -minutes-
for this test.  Mike collected ftrace data that showed that Classic SRCU
was auto-expediting invocations of synchronize_srcu() that found SRCU
completely idle.  This series therefore adds this auto-expedite capability
to Tree SRCU, bringing the performance shortfall to less than ten seconds,
which is a great improvement over the initial shortfall of 15 minutes.

I will continue tuning, but this improvement seemed worth a pull request,
in order to get the base solution some more test time in -tip.  Note that
two of the commits are not strictly necessary, but they are reasonably
simple, they conflict with the three main commits, and the risk of
disentangling them seemed to greatly exceed the risk of pushing them in.
Of course, if you believe that they really should be delayed, please let
me know and I can disentangle them.  They are: (1) c7e88067c1ae ("srcu:
Exact tracking of srcu_data structures containing callbacks") and
(2) 7f6733c3c648 ("srcu: Make rcutorture writer stalls print SRCU GP
state").

These commits are available in the git repository at:

  git://git.kernel.org/pub/scm/linux/kernel/git/paulmck/linux-rcu.git rcu/urgent

for you to fetch changes up to 22607d66bbc3e81140d3bcf08894f4378eb36428:

  srcu: Specify auto-expedite holdoff time (2017-04-26 16:32:17 -0700)

----------------------------------------------------------------
Paul E. McKenney (5):
      srcu: Exact tracking of srcu_data structures containing callbacks
      srcu: Make rcutorture writer stalls print SRCU GP state
      srcu: Expedited grace periods with reduced memory contention
      srcu: Expedite first synchronize_srcu() when idle
      srcu: Specify auto-expedite holdoff time

 Documentation/admin-guide/kernel-parameters.txt |   8 +
 include/linux/srcuclassic.h                     |  14 ++
 include/linux/srcutiny.h                        |  12 ++
 include/linux/srcutree.h                        |  13 +-
 kernel/rcu/rcutorture.c                         |   8 +-
 kernel/rcu/srcutree.c                           | 242 ++++++++++++++++++++----
 kernel/rcu/tree.c                               |  12 +-
 7 files changed, 255 insertions(+), 54 deletions(-)

[toc] | [next] | [standalone]

#1631885

* Paul E. McKenney <paulmck@linux.vnet.ibm.com> wrote:

> Hello, Ingo,
> 
> This series greatly reduces the performance degradation of Tree SRCU
> on a CPU-hotplug stress test.  The effect was not subtle:  Mike Galbraith
> measured Classic SRCU at 55 seconds and Tree SRCU at more than 16 -minutes-
> for this test.  Mike collected ftrace data that showed that Classic SRCU
> was auto-expediting invocations of synchronize_srcu() that found SRCU
> completely idle.  This series therefore adds this auto-expedite capability
> to Tree SRCU, bringing the performance shortfall to less than ten seconds,
> which is a great improvement over the initial shortfall of 15 minutes.
> 
> I will continue tuning, but this improvement seemed worth a pull request,
> in order to get the base solution some more test time in -tip.  Note that
> two of the commits are not strictly necessary, but they are reasonably
> simple, they conflict with the three main commits, and the risk of
> disentangling them seemed to greatly exceed the risk of pushing them in.
> Of course, if you believe that they really should be delayed, please let
> me know and I can disentangle them.  They are: (1) c7e88067c1ae ("srcu:
> Exact tracking of srcu_data structures containing callbacks") and
> (2) 7f6733c3c648 ("srcu: Make rcutorture writer stalls print SRCU GP
> state").
> 
> These commits are available in the git repository at:
> 
>   git://git.kernel.org/pub/scm/linux/kernel/git/paulmck/linux-rcu.git rcu/urgent
> 
> for you to fetch changes up to 22607d66bbc3e81140d3bcf08894f4378eb36428:
> 
>   srcu: Specify auto-expedite holdoff time (2017-04-26 16:32:17 -0700)
> 
> ----------------------------------------------------------------
> Paul E. McKenney (5):
>       srcu: Exact tracking of srcu_data structures containing callbacks
>       srcu: Make rcutorture writer stalls print SRCU GP state
>       srcu: Expedited grace periods with reduced memory contention
>       srcu: Expedite first synchronize_srcu() when idle
>       srcu: Specify auto-expedite holdoff time
> 
>  Documentation/admin-guide/kernel-parameters.txt |   8 +
>  include/linux/srcuclassic.h                     |  14 ++
>  include/linux/srcutiny.h                        |  12 ++
>  include/linux/srcutree.h                        |  13 +-
>  kernel/rcu/rcutorture.c                         |   8 +-
>  kernel/rcu/srcutree.c                           | 242 ++++++++++++++++++++----
>  kernel/rcu/tree.c                               |  12 +-
>  7 files changed, 255 insertions(+), 54 deletions(-)

Pulled, thanks a lot Paul!

	Ingo

[toc] | [prev] | [standalone]

Back to top | Article view | linux.kernel

csiph-web