[lkml]   [2012]   [May]   [3]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
Subject[PATCH v1 0/6] reduce workqueue and timer noise
Timers and work queues both provide a useful way to defer work 
for a later time or a different context. However, when the
timer or work item runs, it interrupts the CPU it is running on.
This is good if it is doing useful work, but it turns out this
is not always the case.

This patch set tries to locate and address code paths where work queues
items and timers are scheduled on CPUs where they have no useful work
to do and adapet them to be more selective.

This includes:

- Introducing helper function to schedule work queue items on a subset
of CPUs in the system.
- Use the helper function to schedule work items to attempt to drain
LRUs only on CPUs where there are LRU pages.
- Stop running the per cpu work item that does per-cpu pages reclaim
and VM statistics on CPUs that did not have any VM activity for the
last second (time frame configurable) and re-start it when VM
activity is detected.
- Fix a bug that prevented the timer code to to not program the
underlying HW timer to fire periodically when no future timer
event exists for a CPU


- The vmstat_update patch was changed to use a scapegoat CPU as
suggested by Christoph Lameter when the patch was previously
discussed in response to Frederic Weisbecker's adaptive tick
patch set.

Also included is a testing only patch, not intdented for mainline,
that turns the clock source watchdog into a config option which
I used while testing the timer code fix change.

The patch was boot tested on 32bit x86 in 8 way SMP and UP VMs.

For you reference, I keep a todo list for these and other noise sources

The git branched can be fetched from the git repo at on the reduce_workqueue_and_timers_noise_v1

Gilad Ben-Yossef (6):
timer: make __next_timer_interrupt explicit about no future event
workqueue: introduce schedule_on_each_cpu_mask
workqueue: introduce schedule_on_each_cpu_cond
mm: make lru_drain selective where it schedules work
mm: make vmstat_update periodic run conditional
x86: make clocksource watchdog configurable (not for mainline)

arch/x86/Kconfig | 9 +++-
include/linux/vmstat.h | 2 +-
include/linux/workqueue.h | 4 ++
kernel/time/clocksource.c | 2 +
kernel/timer.c | 31 ++++++++++-----
kernel/workqueue.c | 73 ++++++++++++++++++++++++++++++----
mm/swap.c | 25 +++++++++++-
mm/vmstat.c | 95 ++++++++++++++++++++++++++++++++++++++-------
8 files changed, 204 insertions(+), 37 deletions(-)

Signed-off-by: Gilad Ben-Yossef <>
CC: Thomas Gleixner <>
CC: Tejun Heo <>
CC: John Stultz <>
CC: Andrew Morton <>
CC: KOSAKI Motohiro <>
CC: Mel Gorman <>
CC: Mike Frysinger <>
CC: David Rientjes <>
CC: Hugh Dickins <>
CC: Minchan Kim <>
CC: Konstantin Khlebnikov <>
CC: Christoph Lameter <>
CC: Chris Metcalf <>
CC: Hakan Akkan <>
CC: Max Krasnyansky <>
CC: Frederic Weisbecker <>

 \ /
  Last update: 2012-05-03 17:41    [W:0.150 / U:2.492 seconds]
©2003-2017 Jasper Spaans. hosted at Digital OceanAdvertise on this site