Impact: scheduling order fix for group scheduling
For each level in the hierarchy, set the buddy to point to the right entity.
Therefore, when we do the hierarchical schedule, we have a fair chance of
ending up where we meant to.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: improve/change/fix wakeup-buddy scheduling
Currently we only have a forward looking buddy, that is, we prefer to
schedule to the task we last woke up, under the presumption that its
going to consume the data we just produced, and therefore will have
cache hot benefits.
This allows co-waking producer/consumer task pairs to run ahead of the
pack for a little while, keeping their cache warm. Without this, we
would interleave all pairs, utterly trashing the cache.
This patch introduces a backward looking buddy, that is, suppose that
in the above scenario, the consumer preempts the producer before it
can go to sleep, we will therefore miss the wakeup from consumer to
producer (its already running, after all), breaking the cycle and
reverting to the cache-trashing interleaved schedule pattern.
The backward buddy will try to schedule back to the task that woke us
up in case the forward buddy is not available, under the assumption
that the last task will be the one with the most cache hot task around
barring current.
This will basically allow a task to continue after it got preempted.
In order to avoid starvation, we allow either buddy to get wakeup_gran
ahead of the pack.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Impact: fix cross-class preemption
Inter-class wakeup preemptions should go on class order.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Since we moved wakeup preemption back to virtual time, it makes sense to move
the buddy stuff back as well. The purpose of the buddy scheduling is to allow
a quickly scheduling pair of tasks to run away from the group as far as a
regular busy task would be allowed under wakeup preemption.
This has the advantage that the pair can ping-pong for a while, enjoying
cache-hotness. Without buddy scheduling other tasks would interleave destroying
the cache.
Also, it saves a word in cfs_rq.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The advantage is that vruntime based wakeup preemption has a better
conceptual model. Here wakeup_gran = 0 means: preempt when 'fair'.
Therefore wakeup_gran is the granularity of unfairness we allow in order
to make progress.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Mysql+oltp and pgsql+oltp peaks are still shifted right. The below puts
the peaks back to 1 client/server pair per core.
Use the avg_overlap information to weaken the sync hint.
Signed-off-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Mike noticed the current min_vruntime tracking can go wrong and skip the
current task. If the only remaining task in the tree is a nice 19 task
with huge vruntime, new tasks will be inserted too far to the right too,
causing some interactibity issues.
min_vruntime can only change due to the leftmost entry disappearing
(dequeue_entity()), or by the leftmost entry being incremented past the
next entry, which elects a new leftmost (__update_curr())
Due to the current entry not being part of the actual tree, we have to
compare the leftmost tree entry with the current entry, and take the
leftmost of these two.
So create a update_min_vruntime() function that takes computes the
leftmost vruntime in the system (either tree of current) and increases
the cfs_rq->min_vruntime if the computed value is larger than the
previously found min_vruntime. And call this from the two sites we've
identified that can change min_vruntime.
Reported-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
* 'sched-fixes-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip:
sched: disable the hrtick for now
sched: revert back to per-rq vruntime
sched: fair scheduler should not resched rt tasks
sched: optimize group load balancer
sched: minor fast-path overhead reduction
sched: fix the wrong mask_len, cleanup
sched: kill unused scheduler decl.
sched: fix the wrong mask_len
sched: only update rq->clock while holding rq->lock
a patch from Henrik Austad did this:
>> Do not declare select_task_rq as part of sched_class when CONFIG_SMP is
>> not set.
Peter observed:
> While a proper cleanup, could you do it by re-arranging the methods so
> as to not create an additional ifdef?
Do not declare select_task_rq and some other methods as part of sched_class
when CONFIG_SMP is not set.
Also gather those methods to avoid CONFIG_SMP mess.
Idea-by: Henrik Austad <henrik.austad@gmail.com>
Signed-off-by: Li Zefan <lizf@cn.fujitsu.com>
Acked-by: Peter Zijlstra <peterz@infradead.org>
Acked-by: Henrik Austad <henrik@austad.us>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Vatsa rightly points out that having the runqueue weight in the vruntime
calculations can cause unfairness in the face of task joins/leaves.
Suppose: dv = dt * rw / w
Then take 10 tasks t_n, each of similar weight. If the first will run 1
then its vruntime will increase by 10. Now, if the next 8 tasks leave after
having run their 1, then the last task will get a vruntime increase of 2
after having run 1.
Which will leave us with 2 tasks of equal weight and equal runtime, of which
one will not be scheduled for 8/2=4 units of time.
Ergo, we cannot do that and must use: dv = dt / w.
This means we cannot have a global vruntime based on effective priority, but
must instead go back to the vruntime per rq model we started out with.
This patch was lightly tested by doing starting while loops on each nice level
and observing their execution time, and a simple group scenario of 1:2:3 pinned
to a single cpu.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
With use of ftrace Steven noticed that some RT tasks got rescheduled due
to sched_fair interaction.
What happens is that we reprogram the hrtick from enqueue/dequeue_fair_task()
because that can change nr_running, and thus a current tasks ideal runtime.
However, its possible the current task isn't a fair_sched_class task, and thus
doesn't have a hrtick set to change.
Fix this by wrapping those hrtick_start_fair() calls in a hrtick_update()
function, which will check for the right conditions.
Reported-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Greetings,
103638d added a bit of avoidable overhead to the fast-path.
Use sysctl_sched_min_granularity instead of sched_slice() to restrict buddy wakeups.
Signed-off-by: Mike Galbraith <efault@gmx.de>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
While looking at the code I wondered why we always do:
sync && avg_overlap < migration_cost
Which is a bit odd, since the overlap test was meant to detect sync wakeups
so using it to specialize sync wakeups doesn't make much sense.
Hence change the code to do:
sync || avg_overlap < migration_cost
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch does following:
o Removes unused variable and argument "rq".
o Optimizes one of the "if" conditions in wake_affine() - i.e. if
"balanced" is true, we need not do rest of the calculations in the
condition.
o If this cpu is same as the previous cpu (on which woken up task
was running when it went to sleep), no need to call wake_affine at all.
Signed-off-by: Amit K Arora <aarora@linux.vnet.ibm.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
cfs_rq->tasks list is used by the load balancer to iterate
over all the tasks. Currently it holds all the entities
(both task and group entities) because of which there is
a need to check for group entities explicitly during load
balancing. This patch changes the cfs_rq->tasks list to
hold only task entities.
Signed-off-by: Bharata B Rao <bharata@linux.vnet.ibm.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
We should set the buddy even though we might already have the
TIF_RESCHED flag set.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
We should not only correct the increment for the initial group, but should
be consistent and do so for all the groups we encounter.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Rework the wakeup preemption to work on real runtime instead of
the virtual runtime. This greatly simplifies the code.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
load_balance_fair() calls rcu_read_lock() but then traverses the list
using the regular list traversal routine. This patch converts the
list traversal to use the _rcu version.
Signed-off-by: Chris Friesen <cfriesen@nortel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Lin Ming reported a 10% OLTP regression against 2.6.27-rc4.
The difference seems to come from different preemption agressiveness,
which affects the cache footprint of the workload and its effective
cache trashing.
Aggresively preempt a task if its avg overlap is very small, this should
avoid the task going to sleep and find it still running when we schedule
back to it - saving a wakeup.
Reported-by: Lin Ming <ming.m.lin@intel.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Overview
This patch reworks the handling of POSIX CPU timers, including the
ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together
with the help of Roland McGrath, the owner and original writer of this code.
The problem we ran into, and the reason for this rework, has to do with using
a profiling timer in a process with a large number of threads. It appears
that the performance of the old implementation of run_posix_cpu_timers() was
at least O(n*3) (where "n" is the number of threads in a process) or worse.
Everything is fine with an increasing number of threads until the time taken
for that routine to run becomes the same as or greater than the tick time, at
which point things degrade rather quickly.
This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF."
Code Changes
This rework corrects the implementation of run_posix_cpu_timers() to make it
run in constant time for a particular machine. (Performance may vary between
one machine and another depending upon whether the kernel is built as single-
or multiprocessor and, in the latter case, depending upon the number of
running processors.) To do this, at each tick we now update fields in
signal_struct as well as task_struct. The run_posix_cpu_timers() function
uses those fields to make its decisions.
We define a new structure, "task_cputime," to contain user, system and
scheduler times and use these in appropriate places:
struct task_cputime {
cputime_t utime;
cputime_t stime;
unsigned long long sum_exec_runtime;
};
This is included in the structure "thread_group_cputime," which is a new
substructure of signal_struct and which varies for uniprocessor versus
multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as
a simple substructure, while for multiprocessor kernels it is a pointer:
struct thread_group_cputime {
struct task_cputime totals;
};
struct thread_group_cputime {
struct task_cputime *totals;
};
We also add a new task_cputime substructure directly to signal_struct, to
cache the earliest expiration of process-wide timers, and task_cputime also
replaces the it_*_expires fields of task_struct (used for earliest expiration
of thread timers). The "thread_group_cputime" structure contains process-wide
timers that are updated via account_user_time() and friends. In the non-SMP
case the structure is a simple aggregator; unfortunately in the SMP case that
simplicity was not achievable due to cache-line contention between CPUs (in
one measured case performance was actually _worse_ on a 16-cpu system than
the same test on a 4-cpu system, due to this contention). For SMP, the
thread_group_cputime counters are maintained as a per-cpu structure allocated
using alloc_percpu(). The timer functions update only the timer field in
the structure corresponding to the running CPU, obtained using per_cpu_ptr().
We define a set of inline functions in sched.h that we use to maintain the
thread_group_cputime structure and hide the differences between UP and SMP
implementations from the rest of the kernel. The thread_group_cputime_init()
function initializes the thread_group_cputime structure for the given task.
The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the
out-of-line function thread_group_cputime_alloc_smp() to allocate and fill
in the per-cpu structures and fields. The thread_group_cputime_free()
function, also a no-op for UP, in SMP frees the per-cpu structures. The
thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls
thread_group_cputime_alloc() if the per-cpu structures haven't yet been
allocated. The thread_group_cputime() function fills the task_cputime
structure it is passed with the contents of the thread_group_cputime fields;
in UP it's that simple but in SMP it must also safely check that tsk->signal
is non-NULL (if it is it just uses the appropriate fields of task_struct) and,
if so, sums the per-cpu values for each online CPU. Finally, the three
functions account_group_user_time(), account_group_system_time() and
account_group_exec_runtime() are used by timer functions to update the
respective fields of the thread_group_cputime structure.
Non-SMP operation is trivial and will not be mentioned further.
The per-cpu structure is always allocated when a task creates its first new
thread, via a call to thread_group_cputime_clone_thread() from copy_signal().
It is freed at process exit via a call to thread_group_cputime_free() from
cleanup_signal().
All functions that formerly summed utime/stime/sum_sched_runtime values from
from all threads in the thread group now use thread_group_cputime() to
snapshot the values in the thread_group_cputime structure or the values in
the task structure itself if the per-cpu structure hasn't been allocated.
Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit.
The run_posix_cpu_timers() function has been split into a fast path and a
slow path; the former safely checks whether there are any expired thread
timers and, if not, just returns, while the slow path does the heavy lifting.
With the dedicated thread group fields, timers are no longer "rebalanced" and
the process_timer_rebalance() function and related code has gone away. All
summing loops are gone and all code that used them now uses the
thread_group_cputime() inline. When process-wide timers are set, the new
task_cputime structure in signal_struct is used to cache the earliest
expiration; this is checked in the fast path.
Performance
The fix appears not to add significant overhead to existing operations. It
generally performs the same as the current code except in two cases, one in
which it performs slightly worse (Case 5 below) and one in which it performs
very significantly better (Case 2 below). Overall it's a wash except in those
two cases.
I've since done somewhat more involved testing on a dual-core Opteron system.
Case 1: With no itimer running, for a test with 100,000 threads, the fixed
kernel took 1428.5 seconds, 513 seconds more than the unfixed system,
all of which was spent in the system. There were twice as many
voluntary context switches with the fix as without it.
Case 2: With an itimer running at .01 second ticks and 4000 threads (the most
an unmodified kernel can handle), the fixed kernel ran the test in
eight percent of the time (5.8 seconds as opposed to 70 seconds) and
had better tick accuracy (.012 seconds per tick as opposed to .023
seconds per tick).
Case 3: A 4000-thread test with an initial timer tick of .01 second and an
interval of 10,000 seconds (i.e. a timer that ticks only once) had
very nearly the same performance in both cases: 6.3 seconds elapsed
for the fixed kernel versus 5.5 seconds for the unfixed kernel.
With fewer threads (eight in these tests), the Case 1 test ran in essentially
the same time on both the modified and unmodified kernels (5.2 seconds versus
5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds
versus 5.4 seconds but again with much better tick accuracy, .013 seconds per
tick versus .025 seconds per tick for the unmodified kernel.
Since the fix affected the rlimit code, I also tested soft and hard CPU limits.
Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer
running), the modified kernel was very slightly favored in that while
it killed the process in 19.997 seconds of CPU time (5.002 seconds of
wall time), only .003 seconds of that was system time, the rest was
user time. The unmodified kernel killed the process in 20.001 seconds
of CPU (5.014 seconds of wall time) of which .016 seconds was system
time. Really, though, the results were too close to call. The results
were essentially the same with no itimer running.
Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds
(where the hard limit would never be reached) and an itimer running,
the modified kernel exhibited worse tick accuracy than the unmodified
kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise,
performance was almost indistinguishable. With no itimer running this
test exhibited virtually identical behavior and times in both cases.
In times past I did some limited performance testing. those results are below.
On a four-cpu Opteron system without this fix, a sixteen-thread test executed
in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On
the same system with the fix, user and elapsed time were about the same, but
system time dropped to 0.007 seconds. Performance with eight, four and one
thread were comparable. Interestingly, the timer ticks with the fix seemed
more accurate: The sixteen-thread test with the fix received 149543 ticks
for 0.024 seconds per tick, while the same test without the fix received 58720
for 0.061 seconds per tick. Both cases were configured for an interval of
0.01 seconds. Again, the other tests were comparable. Each thread in this
test computed the primes up to 25,000,000.
I also did a test with a large number of threads, 100,000 threads, which is
impossible without the fix. In this case each thread computed the primes only
up to 10,000 (to make the runtime manageable). System time dominated, at
1546.968 seconds out of a total 2176.906 seconds (giving a user time of
629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite
accurate. There is obviously no comparable test without the fix.
Signed-off-by: Frank Mayhar <fmayhar@google.com>
Cc: Roland McGrath <roland@redhat.com>
Cc: Alexey Dobriyan <adobriyan@gmail.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The __load_balance_iterator() returns a NULL when there's only one
sched_entity which is a task. It is caused by the following code-path.
/* Skip over entities that are not tasks */
do {
se = list_entry(next, struct sched_entity, group_node);
next = next->next;
} while (next != &cfs_rq->tasks && !entity_is_task(se));
if (next == &cfs_rq->tasks)
return NULL;
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
This will return NULL even when se is a task.
As a side-effect, there was a regression in sched_mc behavior since 2.6.25,
since iter_move_one_task() when it calls load_balance_start_fair(),
would not get any tasks to move!
Fix this by checking if the last entity was a task or not.
Signed-off-by: Gautham R Shenoy <ego@in.ibm.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
- During wake up of a new task, task_new_fair() can do a resched_task()
on the current task. Later in the code path, check_preempt_curr() also ends
up doing the same, which can be avoided. Check if TIF_NEED_RESCHED is
already set for the current task.
- task_new_fair() does a resched_task() on the current task unconditionally.
This can be done only in case when child runs before the parent.
So this is a small speedup.
Signed-off-by: Bharata B Rao <bharata@linux.vnet.ibm.com>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Defer commit 6d299f1b53 to the next release.
Testing of the tip/sched/clock tree revealed a mysql+oltp regression
which bisection eventually traced back to this commit in mainline.
Pertinent test results: Three run sysbench averages, throughput units
in read/write requests/sec.
clients 1 2 4 8 16 32 64
6e0534f 9646 17876 34774 33868 32230 30767 29441
2.6.26.1 9112 17936 34652 33383 31929 30665 29232
6d299f1 9112 14637 28370 33339 32038 30762 29204
Note: subsequent commits hide the majority of this regression until you
apply the clock fixes, at which time it reemerges at full magnitude.
We cannot see anything bad about the change itself so we defer it to the
next release until this problem is fully analysed.
Signed-off-by: Mike Galbraith <efault@gmx.de>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Gregory Haskins <ghaskins@novell.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Benjamin Herrenschmidt reported:
> I get that on ppc64 ...
>
> In file included from kernel/sched.c:1595:
> kernel/sched_fair.c: In function ‘hrtick_start_fair’:
> kernel/sched_fair.c:902: warning: comparison of distinct pointer types lacks a cast
>
> Probably harmless but annoying.
s64 delta = slice - ran;
--> delta = max(10000LL, delta);
Probably ppc64's s64 is long vs long long..
I think hpa was looking at sanitizing all these 64bit types across the
architectures.
Use max_t with an explicit type meanwhile.
Reported-by: Benjamin Herrenschmid <benh@kernel.crashing.org>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
* 'sched/for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/tip/linux-2.6-tip:
sched: hrtick_enabled() should use cpu_active()
sched, x86: clean up hrtick implementation
sched: fix build error, provide partition_sched_domains() unconditionally
sched: fix warning in inc_rt_tasks() to not declare variable 'rq' if it's not needed
cpu hotplug: Make cpu_active_map synchronization dependency clear
cpu hotplug, sched: Introduce cpu_active_map and redo sched domain managment (take 2)
sched: rework of "prioritize non-migratable tasks over migratable ones"
sched: reduce stack size in isolated_cpu_setup()
Revert parts of "ftrace: do not trace scheduler functions"
Fixed up conflicts in include/asm-x86/thread_info.h (due to the
TIF_SINGLESTEP unification vs TIF_HRTICK_RESCHED removal) and
kernel/sched_fair.c (due to cpu_active_map vs for_each_cpu_mask_nr()
introduction).
random uvesafb failures were reported against Gentoo:
http://bugs.gentoo.org/show_bug.cgi?id=222799
and Mihai Moldovan bisected it back to:
> 8f4d37ec07 is first bad commit
> commit 8f4d37ec07
> Author: Peter Zijlstra <a.p.zijlstra@chello.nl>
> Date: Fri Jan 25 21:08:29 2008 +0100
>
> sched: high-res preemption tick
Linus suspected it to be hrtick + vm86 interaction and observed:
> Btw, Peter, Ingo: I think that commit is doing bad things. They aren't
> _incorrect_ per se, but they are definitely bad.
>
> Why?
>
> Using random _TIF_WORK_MASK flags is really impolite for doing
> "scheduling" work. There's a reason that arch/x86/kernel/entry_32.S
> special-cases the _TIF_NEED_RESCHED flag: we don't want to exit out of
> vm86 mode unnecessarily.
>
> See the "work_notifysig_v86" label, and how it does that
> "save_v86_state()" thing etc etc.
Right, I never liked having to fiddle with those TIF flags. Initially I
needed it because the hrtimer base lock could not nest in the rq lock.
That however is fixed these days.
Currently the only reason left to fiddle with the TIF flags is remote
wakeups. We cannot program a remote cpu's hrtimer. I've been thinking
about using the new and improved IPI function call stuff to implement
hrtimer_start_on().
However that does require that smp_call_function_single(.wait=0) works
from interrupt context - /me looks at the latest series from Jens - Yes
that does seem to be supported, good.
Here's a stab at cleaning this stuff up ...
Mihai reported test success as well.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Tested-by: Mihai Moldovan <ionic@ionic.de>
Cc: Michal Januszewski <spock@gentoo.org>
Cc: Antonino Daplas <adaplas@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This is based on Linus' idea of creating cpu_active_map that prevents
scheduler load balancer from migrating tasks to the cpu that is going
down.
It allows us to simplify domain management code and avoid unecessary
domain rebuilds during cpu hotplug event handling.
Please ignore the cpusets part for now. It needs some more work in order
to avoid crazy lock nesting. Although I did simplfy and unify domain
reinitialization logic. We now simply call partition_sched_domains() in
all the cases. This means that we're using exact same code paths as in
cpusets case and hence the test below cover cpusets too.
Cpuset changes to make rebuild_sched_domains() callable from various
contexts are in the separate patch (right next after this one).
This not only boots but also easily handles
while true; do make clean; make -j 8; done
and
while true; do on-off-cpu 1; done
at the same time.
(on-off-cpu 1 simple does echo 0/1 > /sys/.../cpu1/online thing).
Suprisingly the box (dual-core Core2) is quite usable. In fact I'm typing
this on right now in gnome-terminal and things are moving just fine.
Also this is running with most of the debug features enabled (lockdep,
mutex, etc) no BUG_ONs or lockdep complaints so far.
I believe I addressed all of the Dmitry's comments for original Linus'
version. I changed both fair and rt balancer to mask out non-active cpus.
And replaced cpu_is_offline() with !cpu_active() in the main scheduler
code where it made sense (to me).
Signed-off-by: Max Krasnyanskiy <maxk@qualcomm.com>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Acked-by: Gregory Haskins <ghaskins@novell.com>
Cc: dmitry.adamushko@gmail.com
Cc: pj@sgi.com
Signed-off-by: Ingo Molnar <mingo@elte.hu>
We have the notion of tracking process-coupling (a.k.a. buddy-wake) via
the p->se.last_wake / p->se.avg_overlap facilities, but it is only used
for cfs to cfs interactions. There is no reason why an rt to cfs
interaction cannot share in establishing a relationhip in a similar
manner.
Because PREEMPT_RT runs many kernel threads as FIFO priority, we often
times have heavy interaction between RT threads waking CFS applications.
This patch offers a substantial boost (50-60%+) in perfomance under those
circumstances.
Signed-off-by: Gregory Haskins <ghaskins@novell.com>
Cc: npiggin@suse.de
Cc: rostedt@goodmis.org
Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Measurement shows that the difference between cgroup:/ and cgroup:/foo
wake_affine() results is that the latter succeeds significantly more.
Therefore bias the calculations towards failing the test.
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Increase the accuracy of the effective_load values.
Not only consider the current increment (as per the attempted wakeup), but
also consider the delta between when we last adjusted the shares and the
current situation.
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
It was observed these mults can overflow.
Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Currently task_h_load() computes the load of a task and uses that to either
subtract it from the total, or add to it.
However, removing or adding a task need not have any effect on the total load
at all. Imagine adding a task to a group that is local to one cpu - in that
case the total load of that cpu is unaffected.
So properly compute addition/removal:
s_i = S * rw_i / \Sum_j rw_j
s'_i = S * (rw_i + wl) / (\Sum_j rw_j + wg)
then s'_i - s_i gives the change in load.
Where s_i is the shares for cpu i, S the group weight, rw_i the runqueue weight
for that cpu, wl the weight we add (subtract) and wg the weight contribution to
the runqueue.
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
doing the load balance will change cfs_rq->load.weight (that's the whole point)
but since that's part of the scale factor, we'll scale back with a different
amount.
Weight getting smaller would result in an inflated moved_load which causes
it to stop balancing too soon.
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
With hierarchical grouping we can't just compare task weight to rq weight - we
need to scale the weight appropriately.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
While thinking about the previous patch - I realized that using per domain
aggregate load values in load_balance_fair() is wrong. We should use the
load value for that CPU.
By not needing per domain hierarchical load values we don't need to store
per domain aggregate shares, which greatly simplifies all the math.
It basically falls apart in two separate computations:
- per domain update of the shares
- per CPU update of the hierarchical load
Also get rid of the move_group_shares() stuff - just re-compute the shares
again after a successful load balance.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
We only need to know the task_weight of the busiest rq - nothing to do
if there are no tasks there.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The idea was to balance groups until we've reached the global goal, however
Vatsa rightly pointed out that we might never reach that goal this way -
hence take out this logic.
[ the initial rationale for this 'feature' was to promote max concurrency
within a group - it does not however affect fairness ]
Reported-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Keeping the aggregate on the first cpu of the sched domain has two problems:
- it could collide between different sched domains on different cpus
- it could slow things down because of the remote accesses
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Uncouple buddy selection from wakeup granularity.
The initial idea was that buddies could run ahead as far as a normal task
can - do this by measuring a pair 'slice' just as we do for a normal task.
This means we can drop the wakeup_granularity back to 5ms.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Ok, so why are we in this mess, it was:
1/w
but now we mixed that rw in the mix like:
rw/w
rw being \Sum w suggests: fiddling w, we should also fiddle rw, humm?
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
calc_delta_asym() is supposed to do the same as calc_delta_fair() except
linearly shrink the result for negative nice processes - this causes them
to have a smaller preemption threshold so that they are more easily preempted.
The problem is that for task groups se->load.weight is the per cpu share of
the actual task group weight; take that into account.
Also provide a debug switch to disable the asymmetry (which I still don't
like - but it does greatly benefit some workloads)
This would explain the interactivity issues reported against group scheduling.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Cc: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The currently logic inadvertently skips the last task on the run-queue,
resulting in missed balance opportunities.
Signed-off-by: Gregory Haskins <ghaskins@novell.com>
Signed-off-by: David Bahi <dbahi@novell.com>
CC: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Prevent short-running wakers of short-running threads from overloading a single
cpu via wakeup affinity, and wire up disconnected debug option.
Signed-off-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Yanmin Zhang reported:
Comparing with 2.6.25, volanoMark has big regression with kernel 2.6.26-rc1.
It's about 50% on my 8-core stoakley, 16-core tigerton, and Itanium Montecito.
With bisect, I located the following patch:
| 18d95a2832 is first bad commit
| commit 18d95a2832
| Author: Peter Zijlstra <a.p.zijlstra@chello.nl>
| Date: Sat Apr 19 19:45:00 2008 +0200
|
| sched: fair-group: SMP-nice for group scheduling
Revert it so that we get v2.6.25 behavior.
Bisected-by: Yanmin Zhang <yanmin_zhang@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Change references from for_each_cpu_mask to for_each_cpu_mask_nr
where appropriate
Reviewed-by: Paul Jackson <pj@sgi.com>
Reviewed-by: Christoph Lameter <clameter@sgi.com>
Signed-off-by: Mike Travis <travis@sgi.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
The conversion between virtual and real time is as follows:
dvt = rw/w * dt <=> dt = w/rw * dvt
Since we want the fair sleeper granularity to be in real time, we actually
need to do:
dvt = - rw/w * l
This bug could be related to the regression reported by Yanmin Zhang:
| Comparing with kernel 2.6.25, sysbench+mysql(oltp, readonly) has lots
| of regressions with 2.6.26-rc1:
|
| 1) 8-core stoakley: 28%;
| 2) 16-core tigerton: 20%;
| 3) Itanium Montvale: 50%.
Reported-by: "Zhang, Yanmin" <yanmin_zhang@linux.intel.com>
Signed-off-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
this replaces the rq->clock stuff (and possibly cpu_clock()).
- architectures that have an 'imperfect' hardware clock can set
CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
- the 'jiffie' window might be superfulous when we update tick_gtod
before the __update_sched_clock() call in sched_clock_tick()
- cpu_clock() might be implemented as:
sched_clock_cpu(smp_processor_id())
if the accuracy proves good enough - how far can TSC drift in a
single jiffie when considering the filtering and idle hooks?
[ mingo@elte.hu: various fixes and cleanups ]
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Revert debugging commit 7ba2e74ab5.
print_cfs_rq_tasks() can induce live-lock if a task is dequeued
during list traversal.
Signed-off-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
We currently use an optimization to skip the overhead of wake-idle
processing if more than one task is assigned to a run-queue. The
assumption is that the system must already be load-balanced or we
wouldnt be overloaded to begin with.
The problem is that we are looking at rq->nr_running, which may include
RT tasks in addition to CFS tasks. Since the presence of RT tasks
really has no bearing on the balance status of CFS tasks, this throws
the calculation off.
This patch changes the logic to only consider the number of CFS tasks
when making the decision to optimze the wake-idle.
Signed-off-by: Gregory Haskins <ghaskins@novell.com>
CC: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Noticed by sparse:
kernel/sched.c:760:20: warning: symbol 'sched_feat_names' was not declared. Should it be static?
kernel/sched.c:767:5: warning: symbol 'sched_feat_open' was not declared. Should it be static?
kernel/sched_fair.c:845:3: warning: returning void-valued expression
kernel/sched.c:4386:3: warning: returning void-valued expression
Signed-off-by: Harvey Harrison <harvey.harrison@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Normalized sleeper uses calc_delta*() which requires that the rq load is
already updated, so move account_entity_enqueue() before place_entity()
Tested-by: Frans Pop <elendil@planet.nl>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
In order to level the hierarchy, we need to calculate load based on the
root view. That is, each task's load is in the same unit.
A
/ \
B 1
/ \
2 3
To compute 1's load we do:
weight(1)
--------------
rq_weight(A)
To compute 2's load we do:
weight(2) weight(B)
------------ * -----------
rq_weight(B) rw_weight(A)
This yields load fractions in comparable units.
The consequence is that it changes virtual time. We used to have:
time_{i}
vtime_{i} = ------------
weight_{i}
vtime = \Sum vtime_{i} = time / rq_weight.
But with the new way of load calculation we get that vtime equals time.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
De-couple load-balancing from the rb-trees, so that I can change their
organization.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Currently FAIR_GROUP sched grows the scheduler latency outside of
sysctl_sched_latency, invert this so it stays within.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Implement SMP nice support for the full group hierarchy.
On each load-balance action, compile a sched_domain wide view of the full
task_group tree. We compute the domain wide view when walking down the
hierarchy, and readjust the weights when walking back up.
After collecting and readjusting the domain wide view, we try to balance the
tasks within the task_groups. The current approach is a naively balance each
task group until we've moved the targeted amount of load.
Inspired by Srivatsa Vaddsgiri's previous code and Abhishek Chandra's H-SMP
paper.
XXX: there will be some numerical issues due to the limited nature of
SCHED_LOAD_SCALE wrt to representing a task_groups influence on the
total weight. When the tree is deep enough, or the task weight small
enough, we'll run out of bits.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
CC: Abhishek Chandra <chandra@cs.umn.edu>
CC: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
[rebased for sched-devel/latest]
- Add a new cpuset file, having levels:
sched_relax_domain_level
- Modify partition_sched_domains() and build_sched_domains()
to take attributes parameter passed from cpuset.
- Fill newidle_idx for node domains which currently unused but
might be required if sched_relax_domain_level become higher.
- We can change the default level by boot option 'relax_domain_level='.
Signed-off-by: Hidetoshi Seto <seto.hidetoshi@jp.fujitsu.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch allows tasks and groups to exist in the same cfs_rq. With this
change the CFS group scheduling follows a 1/(M+N) model from a 1/(1+N)
fairness model where M tasks and N groups exist at the cfs_rq level.
[a.p.zijlstra@chello.nl: rt bits and assorted fixes]
Signed-off-by: Dhaval Giani <dhaval@linux.vnet.ibm.com>
Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Balbir Singh reported:
> 1:mon> t
> [c0000000e7677da0] c000000000067de0 .sys_sched_yield+0x6c/0xbc
> [c0000000e7677e30] c000000000008748 syscall_exit+0x0/0x40
> --- Exception: c01 (System Call) at 00000400001d09e4
> SP (4000664cb10) is in userspace
> 1:mon> r
> cpu 0x1: Vector: 300 (Data Access) at [c0000000e7677aa0]
> pc: c000000000068e50: .yield_task_fair+0x94/0xc4
> lr: c000000000067de0: .sys_sched_yield+0x6c/0xbc
the check that should have avoided that is:
/*
* Are we the only task in the tree?
*/
if (unlikely(rq->load.weight == curr->se.load.weight))
return;
But I guess that overlooks rt tasks, they also increase the load.
So I guess something like this ought to fix it..
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The wakeup buddy logic didn't use the same wakeup granularity logic as the
wakeup preemption did, this might cause the ->next buddy to be selected past
the point where we would have preempted had the task been a single running
instance.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
revert "sched: fix fair sleepers" (e22ecef1d2),
because it is causing audio skipping, see:
http://bugzilla.kernel.org/show_bug.cgi?id=10428
the patch is correct and the real cause of the skipping is not
understood (tracing makes it go away), but time has run out so we'll
revert it and re-try in 2.6.26.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
TREE_AVG and APPROX_AVG are initial task placement policies that have been
disabled for a long while.. time to remove them.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
CC: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
improve affine wakeups. Maintain the 'overlap' metric based on CFS's
sum_exec_runtime - which means the amount of time a task executes
after it wakes up some other task.
Use the 'overlap' for the wakeup decisions: if the 'overlap' is short,
it means there's strong workload coupling between this task and the
woken up task. If the 'overlap' is large then the workload is decoupled
and the scheduler will move them to separate CPUs more easily.
( Also slightly move the preempt_check within try_to_wake_up() - this has
no effect on functionality but allows 'early wakeups' (for still-on-rq
tasks) to be correctly accounted as well.)
Signed-off-by: Ingo Molnar <mingo@elte.hu>
split out the affine-wakeup bits.
No code changed:
kernel/sched.o:
text data bss dec hex filename
42521 2858 232 45611 b22b sched.o.before
42521 2858 232 45611 b22b sched.o.after
md5:
9d76738f1272aa82f0b7affd2f51df6b sched.o.before.asm
09b31c44e9aff8666f72773dc433e2df sched.o.after.asm
(the md5's changed because stack slots changed and some registers
get scheduled by gcc in a different order - but otherwise the before
and after assembly is instruction for instruction equivalent.)
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Use the existing calc_delta_mine() calculation for sched_slice(). This
saves a divide and simplifies the code because we share it with the
other /cfs_rq->load users.
It also improves code size:
text data bss dec hex filename
42659 2740 144 45543 b1e7 sched.o.before
42093 2740 144 44977 afb1 sched.o.after
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Fair sleepers need to scale their latency target down by runqueue
weight. Otherwise busy systems will gain ever larger sleep bonus.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Currently we schedule to the leftmost task in the runqueue. When the
runtimes are very short because of some server/client ping-pong,
especially in over-saturated workloads, this will cycle through all
tasks trashing the cache.
Reduce cache trashing by keeping dependent tasks together by running
newly woken tasks first. However, by not running the leftmost task first
we could starve tasks because the wakee can gain unlimited runtime.
Therefore we only run the wakee if its within a small
(wakeup_granularity) window of the leftmost task. This preserves
fairness, but does alternate server/client task groups.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Current min_vruntime tracking is incorrect and will cause serious
problems when we don't run the leftmost task for some reason.
min_vruntime does two things; 1) it's used to determine a forward
direction when the u64 vruntime wraps, 2) it's used to track the
leftmost vruntime to position newly enqueued tasks from.
The current logic advances min_vruntime whenever the current task's
vruntime advance. Because the current task may pass the leftmost task
still waiting we're failing the second goal. This causes new tasks to be
placed too far ahead and thus penalizes their runtime.
Fix this by making min_vruntime the min_vruntime of the waiting tasks by
tracking it in enqueue/dequeue, and compare against current's vruntime
to obtain the absolute minimum when placing new tasks.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Kei Tokunaga reported an interactivity problem when moving tasks
between control groups.
Tasks would retain their old vruntime when moved between groups, this
can cause funny lags. Re-set the vruntime on group move to fit within
the new tree.
Reported-by: Kei Tokunaga <tokunaga.keiich@jp.fujitsu.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The following commits cause a number of regressions:
commit 58e2d4ca58
Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Date: Fri Jan 25 21:08:00 2008 +0100
sched: group scheduling, change how cpu load is calculated
commit 6b2d770026
Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Date: Fri Jan 25 21:08:00 2008 +0100
sched: group scheduler, fix fairness of cpu bandwidth allocation for task groups
Namely:
- very frequent wakeups on SMP, reported by PowerTop users.
- cacheline trashing on (large) SMP
- some latencies larger than 500ms
While there is a mergeable patch to fix the latter, the former issues
are not fixable in a manner suitable for .25 (we're at -rc3 now).
Hence we revert them and try again in v2.6.26.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
CC: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Tested-by: Alexey Zaytsev <alexey.zaytsev@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
pick_task_entity() duplicates existing code. This functionality can be
easily obtained using rb_last(). Avoid code duplication by using rb_last().
Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Michel Dänzr has bisected an interactivity problem with
plus-reniced tasks back to this commit:
810e95ccd5 is first bad commit
commit 810e95ccd5
Author: Peter Zijlstra <a.p.zijlstra@chello.nl>
Date: Mon Oct 15 17:00:14 2007 +0200
sched: another wakeup_granularity fix
unit mis-match: wakeup_gran was used against a vruntime
fix this by assymetrically scaling the vtime of positive reniced
tasks.
Bisected-by: Michel Dänzer <michel@tungstengraphics.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The reason why we are getting better wakeup latencies for
!FAIR_USER_SCHED is because of this snippet of code in place_entity():
if (!initial) {
/* sleeps upto a single latency don't count. */
if (sched_feat(NEW_FAIR_SLEEPERS) && entity_is_task(se))
^^^^^^^^^^^^^^^^^^
vruntime -= sysctl_sched_latency;
/* ensure we never gain time by being placed backwards. */
vruntime = max_vruntime(se->vruntime, vruntime);
}
NEW_FAIR_SLEEPERS feature gives credit for sleeping only to tasks and
not group-level entities. With the patch attached, I could see that
wakeup latencies with FAIR_USER_SCHED are restored to the same level as
!FAIR_USER_SCHED.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Right now, the linux kernel (with scheduler statistics enabled) keeps track
of the maximum time a process is waiting to be scheduled. While the maximum
is a very useful metric, tracking average and total is equally useful
(at least for latencytop) to figure out the accumulated effect of scheduler
delays. The accumulated effect is important to judge the performance impact
of scheduler tuning/behavior.
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
print_cfs_stats is callable from interrupt context (sysrq), hence it should
not take mutexes. Change it to use RCU since the task group data is RCU
freed anyway.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
LatencyTOP kernel infrastructure; it measures latencies in the
scheduler and tracks it system wide and per process.
Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Use HR-timers (when available) to deliver an accurate preemption tick.
The regular scheduler tick that runs at 1/HZ can be too coarse when nice
level are used. The fairness system will still keep the cpu utilisation 'fair'
by then delaying the task that got an excessive amount of CPU time but try to
minimize this by delivering preemption points spot-on.
The average frequency of this extra interrupt is sched_latency / nr_latency.
Which need not be higher than 1/HZ, its just that the distribution within the
sched_latency period is important.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Dmitry Adamushko found that the current implementation of the RT
balancing code left out changes to the sched_setscheduler and
rt_mutex_setprio.
This patch addresses this issue by adding methods to the schedule classes
to handle being switched out of (switched_from) and being switched into
(switched_to) a sched_class. Also a method for changing of priorities
is also added (prio_changed).
This patch also removes some duplicate logic between rt_mutex_setprio and
sched_setscheduler.
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Yanmin Zhang noticed a nice optimization:
p = l * nr / nl, nl = l/g -> p = g * nr
which eliminates a do_div() from __sched_period().
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
No need to do a check for 'affine wakeup and passive balancing possibilities'
in select_task_rq_fair() when task_cpu(p) == this_cpu.
I guess, this part got missed upon introduction of per-sched_class
select_task_rq() in try_to_wake_up().
Signed-off-by: Dmitry Adamushko <dmitry.adamushko@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The current wake-up code path tries to determine if it can optimize the
wake-up to "this_cpu" by computing load calculations. The problem is that
these calculations are only relevant to SCHED_OTHER tasks where load is king.
For RT tasks, priority is king. So the load calculation is completely wasted
bandwidth.
Therefore, we create a new sched_class interface to help with
pre-wakeup routing decisions and move the load calculation as a function
of CFS task's class.
Signed-off-by: Gregory Haskins <ghaskins@novell.com>
Signed-off-by: Steven Rostedt <srostedt@redhat.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The current load balancing scheme isn't good enough for precise
group fairness.
For example: on a 8-cpu system, I created 3 groups as under:
a = 8 tasks (cpu.shares = 1024)
b = 4 tasks (cpu.shares = 1024)
c = 3 tasks (cpu.shares = 1024)
a, b and c are task groups that have equal weight. We would expect each
of the groups to receive 33.33% of cpu bandwidth under a fair scheduler.
This is what I get with the latest scheduler git tree:
Signed-off-by: Ingo Molnar <mingo@elte.hu>
--------------------------------------------------------------------------------
Col1 | Col2 | Col3 | Col4
------|---------|-------|-------------------------------------------------------
a | 277.676 | 57.8% | 54.1% 54.1% 54.1% 54.2% 56.7% 62.2% 62.8% 64.5%
b | 116.108 | 24.2% | 47.4% 48.1% 48.7% 49.3%
c | 86.326 | 18.0% | 47.5% 47.9% 48.5%
--------------------------------------------------------------------------------
Explanation of o/p:
Col1 -> Group name
Col2 -> Cumulative execution time (in seconds) received by all tasks of that
group in a 60sec window across 8 cpus
Col3 -> CPU bandwidth received by the group in the 60sec window, expressed in
percentage. Col3 data is derived as:
Col3 = 100 * Col2 / (NR_CPUS * 60)
Col4 -> CPU bandwidth received by each individual task of the group.
Col4 = 100 * cpu_time_recd_by_task / 60
[I can share the test case that produces a similar o/p if reqd]
The deviation from desired group fairness is as below:
a = +24.47%
b = -9.13%
c = -15.33%
which is quite high.
After the patch below is applied, here are the results:
--------------------------------------------------------------------------------
Col1 | Col2 | Col3 | Col4
------|---------|-------|-------------------------------------------------------
a | 163.112 | 34.0% | 33.2% 33.4% 33.5% 33.5% 33.7% 34.4% 34.8% 35.3%
b | 156.220 | 32.5% | 63.3% 64.5% 66.1% 66.5%
c | 160.653 | 33.5% | 85.8% 90.6% 91.4%
--------------------------------------------------------------------------------
Deviation from desired group fairness is as below:
a = +0.67%
b = -0.83%
c = +0.17%
which is far better IMO. Most of other runs have yielded a deviation within
+-2% at the most, which is good.
Why do we see bad (group) fairness with current scheuler?
=========================================================
Currently cpu's weight is just the summation of individual task weights.
This can yield incorrect results. For ex: consider three groups as below
on a 2-cpu system:
CPU0 CPU1
---------------------------
A (10) B(5)
C(5)
---------------------------
Group A has 10 tasks, all on CPU0, Group B and C have 5 tasks each all
of which are on CPU1. Each task has the same weight (NICE_0_LOAD =
1024).
The current scheme would yield a cpu weight of 10240 (10*1024) for each cpu and
the load balancer will think both CPUs are perfectly balanced and won't
move around any tasks. This, however, would yield this bandwidth:
A = 50%
B = 25%
C = 25%
which is not the desired result.
What's changing in the patch?
=============================
- How cpu weights are calculated when CONFIF_FAIR_GROUP_SCHED is
defined (see below)
- API Change
- Two tunables introduced in sysfs (under SCHED_DEBUG) to
control the frequency at which the load balance monitor
thread runs.
The basic change made in this patch is how cpu weight (rq->load.weight) is
calculated. Its now calculated as the summation of group weights on a cpu,
rather than summation of task weights. Weight exerted by a group on a
cpu is dependent on the shares allocated to it and also the number of
tasks the group has on that cpu compared to the total number of
(runnable) tasks the group has in the system.
Let,
W(K,i) = Weight of group K on cpu i
T(K,i) = Task load present in group K's cfs_rq on cpu i
T(K) = Total task load of group K across various cpus
S(K) = Shares allocated to group K
NRCPUS = Number of online cpus in the scheduler domain to
which group K is assigned.
Then,
W(K,i) = S(K) * NRCPUS * T(K,i) / T(K)
A load balance monitor thread is created at bootup, which periodically
runs and adjusts group's weight on each cpu. To avoid its overhead, two
min/max tunables are introduced (under SCHED_DEBUG) to control the rate
at which it runs.
Fixes from: Peter Zijlstra <a.p.zijlstra@chello.nl>
- don't start the load_balance_monitor when there is only a single cpu.
- rename the kthread because its currently longer than TASK_COMM_LEN
Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
This patch changes how the cpu load exerted by fair_sched_class tasks
is calculated. Load exerted by fair_sched_class tasks on a cpu is now
a summation of the group weights, rather than summation of task weights.
Weight exerted by a group on a cpu is dependent on the shares allocated
to it.
This version of patch has a minor impact on code size, but should have
no runtime/functional impact for !CONFIG_FAIR_GROUP_SCHED.
Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Minor bug fixes for the group scheduler:
- Use a mutex to serialize add/remove of task groups and also when
changing shares of a task group. Use the same mutex when printing
cfs_rq debugging stats for various task groups.
- Use list_for_each_entry_rcu in for_each_leaf_cfs_rq macro (when
walking task group list)
Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
measurements by Yanmin Zhang have shown that SCHED_BATCH tasks benefit
if they run the same place_entity() logic as SCHED_OTHER tasks - so
uniformize behavior in this area.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
do more agressive yield for SCHED_BATCH tuned tasks: they are all
about throughput anyway. This allows a gentler migration path for
any apps that relied on stronger yield.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Commit cfb5285660 removed a useful feature for
us, which provided a cpu accounting resource controller. This feature would be
useful if someone wants to group tasks only for accounting purpose and doesnt
really want to exercise any control over their cpu consumption.
The patch below reintroduces the feature. It is based on Paul Menage's
original patch (Commit 62d0df6406), with
these differences:
- Removed load average information. I felt it needs more thought (esp
to deal with SMP and virtualized platforms) and can be added for
2.6.25 after more discussions.
- Convert group cpu usage to be nanosecond accurate (as rest of the cfs
stats are) and invoke cpuacct_charge() from the respective scheduler
classes
- Make accounting scalable on SMP systems by splitting the usage
counter to be per-cpu
- Move the code from kernel/cpu_acct.c to kernel/sched.c (since the
code is not big enough to warrant a new file and also this rightly
needs to live inside the scheduler. Also things like accessing
rq->lock while reading cpu usage becomes easier if the code lived in
kernel/sched.c)
The patch also modifies the cpu controller not to provide the same accounting
information.
Tested-by: Balbir Singh <balbir@linux.vnet.ibm.com>
Tested the patches on top of 2.6.24-rc3. The patches work fine. Ran
some simple tests like cpuspin (spin on the cpu), ran several tasks in
the same group and timed them. Compared their time stamps with
cpuacct.usage.
Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Signed-off-by: Balbir Singh <balbir@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
increase the default minimum granularity some more - this gives us
more performance in aim7 benchmarks.
also correct some comments: we scale with ilog(ncpus) + 1.
Signed-off-by: Zou Nan hai <nanhai.zou@intel.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Sukadev Bhattiprolu reported a kernel crash with control groups.
There are couple of problems discovered by Suka's test:
- The test requires the cgroup filesystem to be mounted with
atleast the cpu and ns options (i.e both namespace and cpu
controllers are active in the same hierarchy).
# mkdir /dev/cpuctl
# mount -t cgroup -ocpu,ns none cpuctl
(or simply)
# mount -t cgroup none cpuctl -> Will activate all controllers
in same hierarchy.
- The test invokes clone() with CLONE_NEWNS set. This causes a a new child
to be created, also a new group (do_fork->copy_namespaces->ns_cgroup_clone->
cgroup_clone) and the child is attached to the new group (cgroup_clone->
attach_task->sched_move_task). At this point in time, the child's scheduler
related fields are uninitialized (including its on_rq field, which it has
inherited from parent). As a result sched_move_task thinks its on
runqueue, when it isn't.
As a solution to this problem, I moved sched_fork() call, which
initializes scheduler related fields on a new task, before
copy_namespaces(). I am not sure though whether moving up will
cause other side-effects. Do you see any issue?
- The second problem exposed by this test is that task_new_fair()
assumes that parent and child will be part of the same group (which
needn't be as this test shows). As a result, cfs_rq->curr can be NULL
for the child.
The solution is to test for curr pointer being NULL in
task_new_fair().
With the patch below, I could run ns_exec() fine w/o a crash.
Reported-by: Sukadev Bhattiprolu <sukadev@us.ibm.com>
Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
clean up the preemption check to not use unnecessary 64-bit
variables. This improves code size:
text data bss dec hex filename
44227 3326 36 47589 b9e5 sched.o.before
44201 3326 36 47563 b9cb sched.o.after
Signed-off-by: Ingo Molnar <mingo@elte.hu>
wakeup preemption fix: do not make it dependent on p->prio.
Preemption purely depends on ->vruntime.
This improves preemption in mixed-nice-level workloads.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
remove PREEMPT_RESTRICT. (this is a separate commit so that any
regression related to the removal itself is bisectable)
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Yanmin Zhang reported an aim7 regression and bisected it down to:
| commit 38ad464d41
| Author: Ingo Molnar <mingo@elte.hu>
| Date: Mon Oct 15 17:00:02 2007 +0200
|
| sched: uniform tunings
|
| use the same defaults on both UP and SMP.
fix this by reintroducing similar SMP tunings again. This resolves
the regression.
(also update the comments to match the ilog2(nr_cpus) tuning effect)
Signed-off-by: Ingo Molnar <mingo@elte.hu>
we lost the sched_min_granularity tunable to a clever optimization
that uses the sched_latency/min_granularity ratio - but the ratio
is quite unintuitive to users and can also crash the kernel if the
ratio is set to 0. So reintroduce the min_granularity tunable,
while keeping the ratio maintained internally.
no functionality changed.
[ mingo@elte.hu: some fixlets. ]
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Add a few comments to place_entity(). No code changed.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
vslice was missing a factor NICE_0_LOAD, as weight is in
weight*NICE_0_LOAD units.
the effect of this bug was larger initial slices and
thus latency-noisier forks.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
fix style of swap() macro in kernel/sched_fair.c.
( this macro should eventually move to a general header, as ext3 uses
a similar construct too. )
Signed-off-by: Ingo Molnar <mingo@elte.hu>
At the moment, a lot of load balancing code that is irrelevant to non
SMP systems gets included during non SMP builds.
This patch addresses this issue and reduces the binary size on non
SMP systems:
text data bss dec hex filename
10983 28 1192 12203 2fab sched.o.before
10739 28 1192 11959 2eb7 sched.o.after
Signed-off-by: Peter Williams <pwil3058@bigpond.net.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
At the moment, balance_tasks() provides low level functionality for both
move_tasks() and move_one_task() (indirectly) via the load_balance()
function (in the sched_class interface) which also provides dual
functionality. This dual functionality complicates the interfaces and
internal mechanisms and makes the run time overhead of operations that
are called with two run queue locks held.
This patch addresses this issue and reduces the overhead of these
operations.
Signed-off-by: Peter Williams <pwil3058@bigpond.net.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Child task may be added on a different cpu that the one on which parent
is running. In which case, task_new_fair() should check whether the new
born task's parent entity should be added as well on the cfs_rq.
Patch below fixes the problem in task_new_fair.
This could fix the put_prev_task_fair() crashes reported.
Reported-by: Kamalesh Babulal <kamalesh@linux.vnet.ibm.com>
Reported-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
reintroduce a simplified version of cache-hot/cold scheduling
affinity. This improves performance with certain SMP workloads,
such as sysbench.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
speed up context-switches a bit by not clearing p->exec_start.
(as a side-effect, this also makes p->exec_start a universal timestamp
available to cache-hot estimations.)
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Prevent wakeup over-scheduling. Once a task has been preempted by a
task of the same or lower priority, it becomes ineligible for repeated
preemption by same until it has been ticked, or slept. Instead, the
task is marked for preemption at the next tick. Tasks of higher
priority still preempt immediately.
Signed-off-by: Mike Galbraith <efault@gmx.de>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Implement feature bit to disable forced preemption. This way
it can be checked whether a workload is overscheduling or not.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
The following patch (sched: disable sleeper_fairness on SCHED_BATCH)
seems to break GROUP_SCHED. Although, it may be 'oops'-less due to the
possibility of 'p' being always a valid address.
Signed-off-by: Dmitry Adamushko <dmitry.adamushko@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
disable sleeper fairness for batch tasks - they are about
batch processing after all.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
unit mis-match: wakeup_gran was used against a vruntime
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
noticed by Peter Zijlstra:
fix: move the CPU check into ->task_new_fair(), this way we
can call place_entity() and get child ->vruntime right at
initial wakeup time.
(without this there can be large latencies)
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
The thing is that __pick_next_entity() must never be called when
first_fair(cfs_rq) == NULL. It wouldn't be a problem, should 'run_node'
be the very first field of 'struct sched_entity' (and it's the second).
The 'nr_running != 0' check is _not_ enough, due to the fact that
'current' is not within the tree. Generic paths are ok (e.g. schedule()
as put_prev_task() is called previously)... I'm more worried about e.g.
migration_call() -> CPU_DEAD_FROZEN -> migrate_dead_tasks()... if
'current' == rq->idle, no problems.. if it's one of the SCHED_NORMAL
tasks (or imagine, some other use-cases in the future -- i.e. we should
not make outer world dependent on internal details of sched_fair class)
-- it may be "Houston, we've got a problem" case.
it's +16 bytes to the ".text". Another variant is to make 'run_node' the
first data member of 'struct sched_entity' but an additional check (se !
= NULL) is still needed in pick_next_entity().
Signed-off-by: Dmitry Adamushko <dmitry.adamushko@gmail.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Make vslice accurate wrt nice levels, and add some comments
while we're at it.
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
mark scheduling classes as const. The speeds up the code
a bit and shrinks it:
text data bss dec hex filename
40027 4018 292 44337 ad31 sched.o.before
40190 3842 292 44324 ad24 sched.o.after
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
There is a possibility that because of task of a group moving from one
cpu to another, it may gain more cpu time that desired. See
http://marc.info/?l=linux-kernel&m=119073197730334 for details.
This is an attempt to fix that problem. Basically it simulates dequeue
of higher level entities as if they are going to sleep. Similarly it
simulate wakeup of higher level entities as if they are waking up from
sleep.
Signed-off-by: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
fix yield bugs due to the current-not-in-rbtree changes: the task is
not in the rbtree so rbtree-removal is a no-no.
[ From: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>: build fix. ]
also, nice code size reduction:
kernel/sched.o:
text data bss dec hex filename
38323 3506 24 41853 a37d sched.o.before
38236 3506 24 41766 a326 sched.o.after
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Dmitry Adamushko <dmitry.adamushko@gmail.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
group scheduler wakeup latency fix: when checking for preemption
we must check cross-group too, not just intra-group.
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Lee Schermerhorn noticed that set_leftmost() contains dead code,
remove this.
Reported-by: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
