q3k
/
linux
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
linux/kernel
History
Nikhil Rao 75dd321d79 sched: Drop group_capacity to 1 only if local group has extra capacity 
When SD_PREFER_SIBLING is set on a sched domain, drop group_capacity to 1
only if the local group has extra capacity. The extra check prevents the case
where you always pull from the heaviest group when it is already under-utilized
(possible with a large weight task outweighs the tasks on the system).

For example, consider a 16-cpu quad-core quad-socket machine with MC and NUMA
scheduling domains. Let's say we spawn 15 nice0 tasks and one nice-15 task,
and each task is running on one core. In this case, we observe the following
events when balancing at the NUMA domain:

- find_busiest_group() will always pick the sched group containing the niced
  task to be the busiest group.
- find_busiest_queue() will then always pick one of the cpus running the
  nice0 task (never picks the cpu with the nice -15 task since
  weighted_cpuload > imbalance).
- The load balancer fails to migrate the task since it is the running task
  and increments sd->nr_balance_failed.
- It repeats the above steps a few more times until sd->nr_balance_failed > 5,
  at which point it kicks off the active load balancer, wakes up the migration
  thread and kicks the nice 0 task off the cpu.

The load balancer doesn't stop until we kick out all nice 0 tasks from
the sched group, leaving you with 3 idle cpus and one cpu running the
nice -15 task.

When balancing at the NUMA domain, we drop sgs.group_capacity to 1 if the child
domain (in this case MC) has SD_PREFER_SIBLING set.  Subsequent load checks are
not relevant because the niced task has a very large weight.

In this patch, we add an extra condition to the "if(prefer_sibling)" check in
update_sd_lb_stats(). We drop the capacity of a group only if the local group
has extra capacity, ie. nr_running < group_capacity. This patch preserves the
original intent of the prefer_siblings check (to spread tasks across the system
in low utilization scenarios) and fixes the case above.

It helps in the following ways:
- In low utilization cases (where nr_tasks << nr_cpus), we still drop
  group_capacity down to 1 if we prefer siblings.
- On very busy systems (where nr_tasks >> nr_cpus), sgs.nr_running will most
  likely be > sgs.group_capacity.
- When balancing large weight tasks, if the local group does not have extra
  capacity, we do not pick the group with the niced task as the busiest group.
  This prevents failed balances, active migration and the under-utilization
  described above.

Signed-off-by: Nikhil Rao <ncrao@google.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <1287173550-30365-5-git-send-email-ncrao@google.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
 		2010-10-18 20:52:19 +02:00
..
debug Merge branch 'core-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip 2010-09-08 11:13:42 -07:00
gcov gcov: fix null-pointer dereference for certain module types 2010-09-09 18:57:23 -07:00
irq irq: Add new IRQ flag IRQF_NO_SUSPEND 2010-07-29 13:24:57 +02:00
power Merge branch 'pm-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/suspend-2.6 2010-09-11 15:50:53 -07:00
time time: Workaround gcc loop optimization that causes 64bit div errors 2010-08-13 12:03:24 -07:00
trace ring-buffer: Fix typo of time extends per page 2010-10-12 12:06:43 -04:00
.gitignore
…
Kconfig.freezer
…
Kconfig.hz
…
Kconfig.locks mutex: Better control mutex adaptive spinning config 2009-12-03 11:50:11 +01:00
Kconfig.preempt
…
Makefile Merge branch 'for-linus' of git://git.infradead.org/users/eparis/notify 2010-08-10 11:39:13 -07:00
acct.c pass a struct path to vfs_statfs 2010-08-09 16:48:42 -04:00
async.c async: use workqueue for worker pool 2010-07-14 11:29:46 +02:00
audit.c Merge branch 'for-linus' of git://git.infradead.org/users/eparis/notify 2010-08-10 11:39:13 -07:00
audit.h Audit: split audit watch Kconfig 2010-07-28 09:58:19 -04:00
audit_tree.c fanotify: use both marks when possible 2010-07-28 10:18:55 -04:00
audit_watch.c Revert "fsnotify: store struct file not struct path" 2010-08-12 14:23:04 -07:00
auditfilter.c audit: do not get and put just to free a watch 2010-07-28 09:58:17 -04:00
auditsc.c vfs: add helpers to get root and pwd 2010-08-11 00:28:20 -04:00
backtracetest.c
…
bounds.c kbuild: move bounds.h to include/generated 2009-12-12 13:08:14 +01:00
capability.c sched: Remove remaining USER_SCHED code 2010-04-02 20:12:00 +02:00
cgroup.c cgroups: fix API thinko 2010-09-09 18:57:23 -07:00
cgroup_freezer.c Freezer / cgroup freezer: Update stale locking comments 2010-05-10 23:18:47 +02:00
compat.c compat: Make compat_alloc_user_space() incorporate the access_ok() 2010-09-14 16:08:45 -07:00
configs.c
…
cpu.c sched: adjust when cpu_active and cpuset configurations are updated during cpu on/offlining 2010-06-08 21:40:36 +02:00
cpuset.c Merge branch 'sched-core-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip 2010-08-06 09:39:22 -07:00
cred.c Add a dummy printk function for the maintenance of unused printks 2010-08-12 09:51:35 -07:00
delayacct.c
…
dma.c
…
early_res.c kmemleak: Add support for NO_BOOTMEM configurations 2010-07-19 11:54:15 +01:00
elfcore.c elf coredump: add extended numbering support 2010-03-06 11:26:46 -08:00
exec_domain.c sys_personality: remove the bogus checks in sys_personality()->__set_personality() path 2010-08-09 20:45:05 -07:00
exit.c Fix unprotected access to task credentials in waitid() 2010-08-17 18:07:43 -07:00
extable.c
…
fork.c rmap: fix walk during fork 2010-09-22 17:22:39 -07:00
freezer.c
…
futex.c futex: futex_find_get_task remove credentails check 2010-06-30 15:43:44 -07:00
futex_compat.c futex: Protect pid lookup in compat code with RCU 2009-12-09 14:22:14 +01:00
groups.c kernel/groups.c: fix integer overflow in groups_search 2010-09-09 18:57:24 -07:00
hrtimer.c gcc-4.6: kernel/*: Fix unused but set warnings 2010-09-05 14:36:58 +02:00
hung_task.c softlockup: Fix hung_task_check_count sysctl 2009-11-27 06:21:57 +01:00
hw_breakpoint.c hw breakpoints: Fix pid namespace bug 2010-09-17 04:42:59 +02:00
itimer.c itimers: Fix racy writes to cpu_itimer fields 2009-11-18 16:32:12 +01:00
kallsyms.c kdb: core for kgdb back end (2 of 2) 2010-05-20 21:04:21 -05:00
kexec.c kexec: return -EFAULT on copy_to_user() failures 2010-08-11 08:59:22 -07:00
kfifo.c kfifo: fix scatterlist usage 2010-10-01 10:50:58 -07:00
kmod.c Make do_execve() take a const filename pointer 2010-08-17 18:07:43 -07:00
kprobes.c kprobes: Move enable/disable_kprobe() out from debugfs code 2010-05-08 18:08:30 +02:00
ksysfs.c sysfs: add struct file* to bin_attr callbacks 2010-05-21 09:37:31 -07:00
kthread.c kthread: implement kthread_data() 2010-06-29 10:07:09 +02:00
latencytop.c include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h 2010-03-30 22:02:32 +09:00
lockdep.c sched_clock: Add local_clock() API and improve documentation 2010-06-09 10:34:49 +02:00
lockdep_internals.h lockdep: No need to disable preemption in debug atomic ops 2010-05-04 05:38:16 +02:00
lockdep_proc.c lockstat: Make lockstat counting per cpu 2010-04-06 00:15:37 +02:00
lockdep_states.h
…
module.c modules: Fix module_bug_list list corruption race 2010-10-05 11:29:27 -07:00
mutex-debug.c
…
mutex-debug.h locking: Implement new raw_spinlock 2009-12-14 23:55:32 +01:00
mutex.c mutex: Fix annotations to include it in kernel-locking docbook 2010-09-03 08:19:51 +02:00
mutex.h
…
notifier.c sched: Use lockdep-based checking on rcu_dereference() 2010-02-25 10:34:26 +01:00
ns_cgroup.c
…
nsproxy.c include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h 2010-03-30 22:02:32 +09:00
padata.c Merge git://git.kernel.org/pub/scm/linux/kernel/git/herbert/crypto-2.6 2010-08-04 15:23:14 -07:00
panic.c lib/bug.c: add oops end marker to WARN implementation 2010-08-11 08:59:22 -07:00
params.c param: locking for kernel parameters 2010-08-11 23:04:20 +09:30
perf_event.c perf: Fix incorrect copy_from_user() usage 2010-10-12 11:45:01 +02:00
pid.c pids: alloc_pidmap: remove the unnecessary boundary checks 2010-08-11 08:59:20 -07:00
pid_namespace.c include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h 2010-03-30 22:02:32 +09:00
pm_qos_params.c PM QoS: Correct pr_debug() misuse and improve parameter checks 2010-09-11 00:53:05 +02:00
posix-cpu-timers.c Merge branch 'writable_limits' of git://decibel.fi.muni.cz/~xslaby/linux 2010-08-10 12:07:51 -07:00
posix-timers.c posix_timer: Move copy_to_user(created_timer_id) down in timer_create() 2010-07-23 15:08:12 +02:00
printk.c gcc-4.6: printk: use stable variable to dump kmsg buffer 2010-08-09 20:45:06 -07:00
profile.c numa: in-kernel profiling: use cpu_to_mem() for per cpu allocations 2010-05-27 09:12:57 -07:00
ptrace.c ptrace: optimize exit_ptrace() for the likely case 2010-08-11 08:59:19 -07:00
range.c kernel/range: remove unused definition of ARRAY_SIZE() 2010-08-09 20:45:06 -07:00
rcupdate.c tree/tiny rcu: Add debug RCU head objects 2010-06-14 16:37:26 -07:00
rcutiny.c tree/tiny rcu: Add debug RCU head objects 2010-06-14 16:37:26 -07:00
rcutiny_plugin.h rcu: slim down rcutiny by removing rcu_scheduler_active and friends 2010-05-10 11:08:34 -07:00
rcutorture.c sched_clock: Add local_clock() API and improve documentation 2010-06-09 10:34:49 +02:00
rcutree.c tree/tiny rcu: Add debug RCU head objects 2010-06-14 16:37:26 -07:00
rcutree.h rcu: reduce the number of spurious RCU_SOFTIRQ invocations 2010-05-10 11:08:35 -07:00
rcutree_plugin.h rcu: remove all rcu head initializations, except on_stack initializations 2010-05-11 16:10:47 -07:00
rcutree_trace.c rcu: reduce the number of spurious RCU_SOFTIRQ invocations 2010-05-10 11:08:35 -07:00
relay.c kernel/: convert cpu notifier to return encapsulate errno value 2010-05-27 09:12:48 -07:00
res_counter.c include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h 2010-03-30 22:02:32 +09:00
resource.c resource: shared I/O region support 2010-05-11 12:01:10 -07:00
rtmutex-debug.c sched: Convert pi_lock to raw_spinlock 2009-12-14 23:55:33 +01:00
rtmutex-debug.h
…
rtmutex-tester.c
…
rtmutex.c rtmutes: Convert rtmutex.lock to raw_spinlock 2009-12-14 23:55:33 +01:00
rtmutex.h
…
rtmutex_common.h
…
rwsem.c
…
sched.c sched: Do not consider SCHED_IDLE tasks to be cache hot 2010-10-18 20:52:15 +02:00
sched_clock.c sched_clock: Add local_clock() API and improve documentation 2010-06-09 10:34:49 +02:00
sched_cpupri.c sched: No need for bootmem special cases 2010-07-17 12:06:22 +02:00
sched_cpupri.h sched: No need for bootmem special cases 2010-07-17 12:06:22 +02:00
sched_debug.c sched: Use correct macro to display sched_child_runs_first in /proc/sched_debug 2010-07-21 21:46:12 +02:00
sched_fair.c sched: Drop group_capacity to 1 only if local group has extra capacity 2010-10-18 20:52:19 +02:00
sched_features.h sched: Remove ASYM_GRAN feature 2010-03-11 18:32:53 +01:00
sched_idletask.c sched: Cure load average vs NO_HZ woes 2010-04-23 11:02:02 +02:00
sched_rt.c sched: Unindent labels 2010-10-18 18:41:56 +02:00
sched_stats.h sched: Remove the obsolete exit_state/signal hacks 2010-06-18 10:46:56 +02:00
sched_stoptask.c sched: Create special class for stop/migrate work 2010-10-18 18:41:58 +02:00
seccomp.c
…
semaphore.c
…
signal.c HWPOISON: Copy si_addr_lsb to user 2010-10-07 09:41:25 +02:00
smp.c generic-ipi: Fix deadlock in __smp_call_function_single 2010-09-10 16:48:40 +02:00
softirq.c kernel/: fix BUG_ON checks for cpu notifier callbacks direct call 2010-06-04 15:21:45 -07:00
spinlock.c locking: Cleanup the name space completely 2009-12-14 23:55:33 +01:00
srcu.c include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h 2010-03-30 22:02:32 +09:00
stacktrace.c
…
stop_machine.c sched: Create special class for stop/migrate work 2010-10-18 18:41:58 +02:00
sys.c pid: make setpgid() system call use RCU read-side critical section 2010-08-31 17:00:18 -07:00
sys_ni.c fanotify: sys_fanotify_mark declartion 2010-07-28 09:58:55 -04:00
sysctl.c sysctl: fix min/max handling in __do_proc_doulongvec_minmax() 2010-10-07 13:31:21 -07:00
sysctl_binary.c sysctl: don't use own implementation of hex_to_bin() 2010-05-25 08:07:05 -07:00
sysctl_check.c ipv4 05/05: add sysctl to accept packets with local source addresses 2009-12-03 12:14:38 -08:00
taskstats.c include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit slab.h inclusion from percpu.h 2010-03-30 22:02:32 +09:00
test_kprobes.c
…
time.c time: Kill off CONFIG_GENERIC_TIME 2010-07-27 12:40:54 +02:00
timeconst.pl
…
timer.c kernel/timer.c: fix kernel-doc function parameter warning 2010-08-10 15:33:09 -07:00
tracepoint.c tracing: Let tracepoints have data passed to tracepoint callbacks 2010-05-14 09:50:34 -04:00
tsacct.c mm: clean up mm_counter 2010-03-06 11:26:23 -08:00
uid16.c
…
up.c
…
user-return-notifier.c core: Clean up user return notifers use of per_cpu 2009-12-02 10:22:59 +01:00
user.c sched: Remove a stale comment 2010-05-10 08:48:39 +02:00
user_namespace.c user_ns: Introduce user_nsmap_uid and user_ns_map_gid. 2010-06-16 14:55:34 -07:00
utsname.c
…
utsname_sysctl.c
…
wait.c
…
watchdog.c lockup_detector: Sync touch_*_watchdog back to old semantics 2010-09-01 10:02:28 +02:00
workqueue.c workqueue: add documentation 2010-09-13 10:26:52 +02:00
workqueue_sched.h workqueue: implement concurrency managed dynamic worker pool 2010-06-29 10:07:14 +02:00