linux/kernel/sched_fair.c

1446 lines
35 KiB
C
Raw Normal View History

/*
* Completely Fair Scheduling (CFS) Class (SCHED_NORMAL/SCHED_BATCH)
*
* Copyright (C) 2007 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
*
* Interactivity improvements by Mike Galbraith
* (C) 2007 Mike Galbraith <efault@gmx.de>
*
* Various enhancements by Dmitry Adamushko.
* (C) 2007 Dmitry Adamushko <dmitry.adamushko@gmail.com>
*
* Group scheduling enhancements by Srivatsa Vaddagiri
* Copyright IBM Corporation, 2007
* Author: Srivatsa Vaddagiri <vatsa@linux.vnet.ibm.com>
*
* Scaled math optimizations by Thomas Gleixner
* Copyright (C) 2007, Thomas Gleixner <tglx@linutronix.de>
*
* Adaptive scheduling granularity, math enhancements by Peter Zijlstra
* Copyright (C) 2007 Red Hat, Inc., Peter Zijlstra <pzijlstr@redhat.com>
*/
#include <linux/latencytop.h>
/*
* Targeted preemption latency for CPU-bound tasks:
* (default: 20ms * (1 + ilog(ncpus)), units: nanoseconds)
*
* NOTE: this latency value is not the same as the concept of
* 'timeslice length' - timeslices in CFS are of variable length
* and have no persistent notion like in traditional, time-slice
* based scheduling concepts.
*
* (to see the precise effective timeslice length of your workload,
* run vmstat and monitor the context-switches (cs) field)
*/
unsigned int sysctl_sched_latency = 20000000ULL;
/*
* Minimal preemption granularity for CPU-bound tasks:
* (default: 4 msec * (1 + ilog(ncpus)), units: nanoseconds)
*/
unsigned int sysctl_sched_min_granularity = 4000000ULL;
/*
* is kept at sysctl_sched_latency / sysctl_sched_min_granularity
*/
static unsigned int sched_nr_latency = 5;
/*
* After fork, child runs first. (default) If set to 0 then
* parent will (try to) run first.
*/
const_debug unsigned int sysctl_sched_child_runs_first = 1;
/*
* sys_sched_yield() compat mode
*
* This option switches the agressive yield implementation of the
* old scheduler back on.
*/
unsigned int __read_mostly sysctl_sched_compat_yield;
/*
* SCHED_BATCH wake-up granularity.
* (default: 10 msec * (1 + ilog(ncpus)), units: nanoseconds)
*
* This option delays the preemption effects of decoupled workloads
* and reduces their over-scheduling. Synchronous workloads will still
* have immediate wakeup/sleep latencies.
*/
unsigned int sysctl_sched_batch_wakeup_granularity = 10000000UL;
/*
* SCHED_OTHER wake-up granularity.
* (default: 10 msec * (1 + ilog(ncpus)), units: nanoseconds)
*
* This option delays the preemption effects of decoupled workloads
* and reduces their over-scheduling. Synchronous workloads will still
* have immediate wakeup/sleep latencies.
*/
unsigned int sysctl_sched_wakeup_granularity = 10000000UL;
const_debug unsigned int sysctl_sched_migration_cost = 500000UL;
/**************************************************************
* CFS operations on generic schedulable entities:
*/
#ifdef CONFIG_FAIR_GROUP_SCHED
/* cpu runqueue to which this cfs_rq is attached */
static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
{
return cfs_rq->rq;
}
/* An entity is a task if it doesn't "own" a runqueue */
#define entity_is_task(se) (!se->my_q)
#else /* CONFIG_FAIR_GROUP_SCHED */
static inline struct rq *rq_of(struct cfs_rq *cfs_rq)
{
return container_of(cfs_rq, struct rq, cfs);
}
#define entity_is_task(se) 1
#endif /* CONFIG_FAIR_GROUP_SCHED */
static inline struct task_struct *task_of(struct sched_entity *se)
{
return container_of(se, struct task_struct, se);
}
/**************************************************************
* Scheduling class tree data structure manipulation methods:
*/
static inline u64 max_vruntime(u64 min_vruntime, u64 vruntime)
{
s64 delta = (s64)(vruntime - min_vruntime);
if (delta > 0)
min_vruntime = vruntime;
return min_vruntime;
}
static inline u64 min_vruntime(u64 min_vruntime, u64 vruntime)
{
s64 delta = (s64)(vruntime - min_vruntime);
if (delta < 0)
min_vruntime = vruntime;
return min_vruntime;
}
static inline s64 entity_key(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
return se->vruntime - cfs_rq->min_vruntime;
}
/*
* Enqueue an entity into the rb-tree:
*/
static void __enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
struct rb_node *parent = NULL;
struct sched_entity *entry;
s64 key = entity_key(cfs_rq, se);
int leftmost = 1;
/*
* Find the right place in the rbtree:
*/
while (*link) {
parent = *link;
entry = rb_entry(parent, struct sched_entity, run_node);
/*
* We dont care about collisions. Nodes with
* the same key stay together.
*/
if (key < entity_key(cfs_rq, entry)) {
link = &parent->rb_left;
} else {
link = &parent->rb_right;
leftmost = 0;
}
}
/*
* Maintain a cache of leftmost tree entries (it is frequently
* used):
*/
if (leftmost)
cfs_rq->rb_leftmost = &se->run_node;
rb_link_node(&se->run_node, parent, link);
rb_insert_color(&se->run_node, &cfs_rq->tasks_timeline);
}
static void __dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
if (cfs_rq->rb_leftmost == &se->run_node)
cfs_rq->rb_leftmost = rb_next(&se->run_node);
rb_erase(&se->run_node, &cfs_rq->tasks_timeline);
}
static inline struct rb_node *first_fair(struct cfs_rq *cfs_rq)
{
return cfs_rq->rb_leftmost;
}
static struct sched_entity *__pick_next_entity(struct cfs_rq *cfs_rq)
{
return rb_entry(first_fair(cfs_rq), struct sched_entity, run_node);
}
static inline struct sched_entity *__pick_last_entity(struct cfs_rq *cfs_rq)
{
struct rb_node **link = &cfs_rq->tasks_timeline.rb_node;
struct sched_entity *se = NULL;
struct rb_node *parent;
while (*link) {
parent = *link;
se = rb_entry(parent, struct sched_entity, run_node);
link = &parent->rb_right;
}
return se;
}
/**************************************************************
* Scheduling class statistics methods:
*/
#ifdef CONFIG_SCHED_DEBUG
int sched_nr_latency_handler(struct ctl_table *table, int write,
struct file *filp, void __user *buffer, size_t *lenp,
loff_t *ppos)
{
int ret = proc_dointvec_minmax(table, write, filp, buffer, lenp, ppos);
if (ret || !write)
return ret;
sched_nr_latency = DIV_ROUND_UP(sysctl_sched_latency,
sysctl_sched_min_granularity);
return 0;
}
#endif
/*
* The idea is to set a period in which each task runs once.
*
* When there are too many tasks (sysctl_sched_nr_latency) we have to stretch
* this period because otherwise the slices get too small.
*
* p = (nr <= nl) ? l : l*nr/nl
*/
static u64 __sched_period(unsigned long nr_running)
{
u64 period = sysctl_sched_latency;
unsigned long nr_latency = sched_nr_latency;
if (unlikely(nr_running > nr_latency)) {
period = sysctl_sched_min_granularity;
period *= nr_running;
}
return period;
}
/*
* We calculate the wall-time slice from the period by taking a part
* proportional to the weight.
*
* s = p*w/rw
*/
static u64 sched_slice(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
u64 slice = __sched_period(cfs_rq->nr_running);
slice *= se->load.weight;
do_div(slice, cfs_rq->load.weight);
return slice;
}
/*
* We calculate the vruntime slice.
*
* vs = s/w = p/rw
*/
static u64 __sched_vslice(unsigned long rq_weight, unsigned long nr_running)
{
u64 vslice = __sched_period(nr_running);
vslice *= NICE_0_LOAD;
do_div(vslice, rq_weight);
return vslice;
}
static u64 sched_vslice(struct cfs_rq *cfs_rq)
{
return __sched_vslice(cfs_rq->load.weight, cfs_rq->nr_running);
}
static u64 sched_vslice_add(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
return __sched_vslice(cfs_rq->load.weight + se->load.weight,
cfs_rq->nr_running + 1);
}
/*
* Update the current task's runtime statistics. Skip current tasks that
* are not in our scheduling class.
*/
static inline void
__update_curr(struct cfs_rq *cfs_rq, struct sched_entity *curr,
unsigned long delta_exec)
{
unsigned long delta_exec_weighted;
u64 vruntime;
schedstat_set(curr->exec_max, max((u64)delta_exec, curr->exec_max));
curr->sum_exec_runtime += delta_exec;
schedstat_add(cfs_rq, exec_clock, delta_exec);
delta_exec_weighted = delta_exec;
if (unlikely(curr->load.weight != NICE_0_LOAD)) {
delta_exec_weighted = calc_delta_fair(delta_exec_weighted,
&curr->load);
}
curr->vruntime += delta_exec_weighted;
/*
* maintain cfs_rq->min_vruntime to be a monotonic increasing
* value tracking the leftmost vruntime in the tree.
*/
if (first_fair(cfs_rq)) {
vruntime = min_vruntime(curr->vruntime,
__pick_next_entity(cfs_rq)->vruntime);
} else
vruntime = curr->vruntime;
cfs_rq->min_vruntime =
max_vruntime(cfs_rq->min_vruntime, vruntime);
}
static void update_curr(struct cfs_rq *cfs_rq)
{
struct sched_entity *curr = cfs_rq->curr;
u64 now = rq_of(cfs_rq)->clock;
unsigned long delta_exec;
if (unlikely(!curr))
return;
/*
* Get the amount of time the current task was running
* since the last time we changed load (this cannot
* overflow on 32 bits):
*/
delta_exec = (unsigned long)(now - curr->exec_start);
__update_curr(cfs_rq, curr, delta_exec);
curr->exec_start = now;
sched: cpu accounting controller (V2) Commit cfb5285660aad4931b2ebbfa902ea48a37dfffa1 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 62d0df64065e7c135d0002f069444fbdfc64768f), 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>
2007-12-02 19:04:49 +00:00
if (entity_is_task(curr)) {
struct task_struct *curtask = task_of(curr);
cpuacct_charge(curtask, delta_exec);
}
}
static inline void
update_stats_wait_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
schedstat_set(se->wait_start, rq_of(cfs_rq)->clock);
}
/*
* Task is being enqueued - update stats:
*/
static void update_stats_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
/*
* Are we enqueueing a waiting task? (for current tasks
* a dequeue/enqueue event is a NOP)
*/
if (se != cfs_rq->curr)
update_stats_wait_start(cfs_rq, se);
}
static void
update_stats_wait_end(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
schedstat_set(se->wait_max, max(se->wait_max,
rq_of(cfs_rq)->clock - se->wait_start));
schedstat_set(se->wait_count, se->wait_count + 1);
schedstat_set(se->wait_sum, se->wait_sum +
rq_of(cfs_rq)->clock - se->wait_start);
schedstat_set(se->wait_start, 0);
}
static inline void
update_stats_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
/*
* Mark the end of the wait period if dequeueing a
* waiting task:
*/
if (se != cfs_rq->curr)
update_stats_wait_end(cfs_rq, se);
}
/*
* We are picking a new current task - update its stats:
*/
static inline void
update_stats_curr_start(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
/*
* We are starting a new run period:
*/
se->exec_start = rq_of(cfs_rq)->clock;
}
/**************************************************
* Scheduling class queueing methods:
*/
static void
account_entity_enqueue(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
update_load_add(&cfs_rq->load, se->load.weight);
cfs_rq->nr_running++;
se->on_rq = 1;
}
static void
account_entity_dequeue(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
update_load_sub(&cfs_rq->load, se->load.weight);
cfs_rq->nr_running--;
se->on_rq = 0;
}
static void enqueue_sleeper(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
#ifdef CONFIG_SCHEDSTATS
if (se->sleep_start) {
u64 delta = rq_of(cfs_rq)->clock - se->sleep_start;
struct task_struct *tsk = task_of(se);
if ((s64)delta < 0)
delta = 0;
if (unlikely(delta > se->sleep_max))
se->sleep_max = delta;
se->sleep_start = 0;
se->sum_sleep_runtime += delta;
account_scheduler_latency(tsk, delta >> 10, 1);
}
if (se->block_start) {
u64 delta = rq_of(cfs_rq)->clock - se->block_start;
struct task_struct *tsk = task_of(se);
if ((s64)delta < 0)
delta = 0;
if (unlikely(delta > se->block_max))
se->block_max = delta;
se->block_start = 0;
se->sum_sleep_runtime += delta;
/*
* Blocking time is in units of nanosecs, so shift by 20 to
* get a milliseconds-range estimation of the amount of
* time that the task spent sleeping:
*/
if (unlikely(prof_on == SLEEP_PROFILING)) {
profile_hits(SLEEP_PROFILING, (void *)get_wchan(tsk),
delta >> 20);
}
account_scheduler_latency(tsk, delta >> 10, 0);
}
#endif
}
static void check_spread(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
#ifdef CONFIG_SCHED_DEBUG
s64 d = se->vruntime - cfs_rq->min_vruntime;
if (d < 0)
d = -d;
if (d > 3*sysctl_sched_latency)
schedstat_inc(cfs_rq, nr_spread_over);
#endif
}
static void
place_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int initial)
{
u64 vruntime;
vruntime = cfs_rq->min_vruntime;
if (sched_feat(TREE_AVG)) {
struct sched_entity *last = __pick_last_entity(cfs_rq);
if (last) {
vruntime += last->vruntime;
vruntime >>= 1;
}
} else if (sched_feat(APPROX_AVG) && cfs_rq->nr_running)
vruntime += sched_vslice(cfs_rq)/2;
/*
* The 'current' period is already promised to the current tasks,
* however the extra weight of the new task will slow them down a
* little, place the new task so that it fits in the slot that
* stays open at the end.
*/
if (initial && sched_feat(START_DEBIT))
vruntime += sched_vslice_add(cfs_rq, se);
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);
}
se->vruntime = vruntime;
}
static void
enqueue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int wakeup)
{
/*
* Update run-time statistics of the 'current'.
*/
update_curr(cfs_rq);
if (wakeup) {
place_entity(cfs_rq, se, 0);
enqueue_sleeper(cfs_rq, se);
}
update_stats_enqueue(cfs_rq, se);
check_spread(cfs_rq, se);
if (se != cfs_rq->curr)
__enqueue_entity(cfs_rq, se);
account_entity_enqueue(cfs_rq, se);
}
static void
dequeue_entity(struct cfs_rq *cfs_rq, struct sched_entity *se, int sleep)
{
/*
* Update run-time statistics of the 'current'.
*/
update_curr(cfs_rq);
update_stats_dequeue(cfs_rq, se);
if (sleep) {
#ifdef CONFIG_SCHEDSTATS
if (entity_is_task(se)) {
struct task_struct *tsk = task_of(se);
if (tsk->state & TASK_INTERRUPTIBLE)
se->sleep_start = rq_of(cfs_rq)->clock;
if (tsk->state & TASK_UNINTERRUPTIBLE)
se->block_start = rq_of(cfs_rq)->clock;
}
#endif
}
if (se != cfs_rq->curr)
__dequeue_entity(cfs_rq, se);
account_entity_dequeue(cfs_rq, se);
}
/*
* Preempt the current task with a newly woken task if needed:
*/
static void
check_preempt_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr)
{
unsigned long ideal_runtime, delta_exec;
ideal_runtime = sched_slice(cfs_rq, curr);
delta_exec = curr->sum_exec_runtime - curr->prev_sum_exec_runtime;
if (delta_exec > ideal_runtime)
resched_task(rq_of(cfs_rq)->curr);
}
static void
set_next_entity(struct cfs_rq *cfs_rq, struct sched_entity *se)
{
/* 'current' is not kept within the tree. */
if (se->on_rq) {
/*
* Any task has to be enqueued before it get to execute on
* a CPU. So account for the time it spent waiting on the
* runqueue.
*/
update_stats_wait_end(cfs_rq, se);
__dequeue_entity(cfs_rq, se);
}
update_stats_curr_start(cfs_rq, se);
cfs_rq->curr = se;
#ifdef CONFIG_SCHEDSTATS
/*
* Track our maximum slice length, if the CPU's load is at
* least twice that of our own weight (i.e. dont track it
* when there are only lesser-weight tasks around):
*/
if (rq_of(cfs_rq)->load.weight >= 2*se->load.weight) {
se->slice_max = max(se->slice_max,
se->sum_exec_runtime - se->prev_sum_exec_runtime);
}
#endif
se->prev_sum_exec_runtime = se->sum_exec_runtime;
}
static struct sched_entity *pick_next_entity(struct cfs_rq *cfs_rq)
{
struct sched_entity *se = NULL;
if (first_fair(cfs_rq)) {
se = __pick_next_entity(cfs_rq);
set_next_entity(cfs_rq, se);
}
return se;
}
static void put_prev_entity(struct cfs_rq *cfs_rq, struct sched_entity *prev)
{
/*
* If still on the runqueue then deactivate_task()
* was not called and update_curr() has to be done:
*/
if (prev->on_rq)
update_curr(cfs_rq);
check_spread(cfs_rq, prev);
if (prev->on_rq) {
update_stats_wait_start(cfs_rq, prev);
/* Put 'current' back into the tree. */
__enqueue_entity(cfs_rq, prev);
}
cfs_rq->curr = NULL;
}
static void
entity_tick(struct cfs_rq *cfs_rq, struct sched_entity *curr, int queued)
{
/*
* Update run-time statistics of the 'current'.
*/
update_curr(cfs_rq);
#ifdef CONFIG_SCHED_HRTICK
/*
* queued ticks are scheduled to match the slice, so don't bother
* validating it and just reschedule.
*/
if (queued)
return resched_task(rq_of(cfs_rq)->curr);
/*
* don't let the period tick interfere with the hrtick preemption
*/
if (!sched_feat(DOUBLE_TICK) &&
hrtimer_active(&rq_of(cfs_rq)->hrtick_timer))
return;
#endif
if (cfs_rq->nr_running > 1 || !sched_feat(WAKEUP_PREEMPT))
check_preempt_tick(cfs_rq, curr);
}
/**************************************************
* CFS operations on tasks:
*/
#ifdef CONFIG_FAIR_GROUP_SCHED
/* Walk up scheduling entities hierarchy */
#define for_each_sched_entity(se) \
for (; se; se = se->parent)
static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
{
return p->se.cfs_rq;
}
/* runqueue on which this entity is (to be) queued */
static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
{
return se->cfs_rq;
}
/* runqueue "owned" by this group */
static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
{
return grp->my_q;
}
/* Given a group's cfs_rq on one cpu, return its corresponding cfs_rq on
* another cpu ('this_cpu')
*/
static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
{
return cfs_rq->tg->cfs_rq[this_cpu];
}
/* Iterate thr' all leaf cfs_rq's on a runqueue */
#define for_each_leaf_cfs_rq(rq, cfs_rq) \
list_for_each_entry_rcu(cfs_rq, &rq->leaf_cfs_rq_list, leaf_cfs_rq_list)
/* Do the two (enqueued) entities belong to the same group ? */
static inline int
is_same_group(struct sched_entity *se, struct sched_entity *pse)
{
if (se->cfs_rq == pse->cfs_rq)
return 1;
return 0;
}
static inline struct sched_entity *parent_entity(struct sched_entity *se)
{
return se->parent;
}
sched: group scheduler, fix fairness of cpu bandwidth allocation for task groups 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>
2008-01-25 20:08:00 +00:00
#define GROUP_IMBALANCE_PCT 20
#else /* CONFIG_FAIR_GROUP_SCHED */
#define for_each_sched_entity(se) \
for (; se; se = NULL)
static inline struct cfs_rq *task_cfs_rq(struct task_struct *p)
{
return &task_rq(p)->cfs;
}
static inline struct cfs_rq *cfs_rq_of(struct sched_entity *se)
{
struct task_struct *p = task_of(se);
struct rq *rq = task_rq(p);
return &rq->cfs;
}
/* runqueue "owned" by this group */
static inline struct cfs_rq *group_cfs_rq(struct sched_entity *grp)
{
return NULL;
}
static inline struct cfs_rq *cpu_cfs_rq(struct cfs_rq *cfs_rq, int this_cpu)
{
return &cpu_rq(this_cpu)->cfs;
}
#define for_each_leaf_cfs_rq(rq, cfs_rq) \
for (cfs_rq = &rq->cfs; cfs_rq; cfs_rq = NULL)
static inline int
is_same_group(struct sched_entity *se, struct sched_entity *pse)
{
return 1;
}
static inline struct sched_entity *parent_entity(struct sched_entity *se)
{
return NULL;
}
#endif /* CONFIG_FAIR_GROUP_SCHED */
#ifdef CONFIG_SCHED_HRTICK
static void hrtick_start_fair(struct rq *rq, struct task_struct *p)
{
int requeue = rq->curr == p;
struct sched_entity *se = &p->se;
struct cfs_rq *cfs_rq = cfs_rq_of(se);
WARN_ON(task_rq(p) != rq);
if (hrtick_enabled(rq) && cfs_rq->nr_running > 1) {
u64 slice = sched_slice(cfs_rq, se);
u64 ran = se->sum_exec_runtime - se->prev_sum_exec_runtime;
s64 delta = slice - ran;
if (delta < 0) {
if (rq->curr == p)
resched_task(p);
return;
}
/*
* Don't schedule slices shorter than 10000ns, that just
* doesn't make sense. Rely on vruntime for fairness.
*/
if (!requeue)
delta = max(10000LL, delta);
hrtick_start(rq, delta, requeue);
}
}
#else
static inline void
hrtick_start_fair(struct rq *rq, struct task_struct *p)
{
}
#endif
/*
* The enqueue_task method is called before nr_running is
* increased. Here we update the fair scheduling stats and
* then put the task into the rbtree:
*/
static void enqueue_task_fair(struct rq *rq, struct task_struct *p, int wakeup)
{
struct cfs_rq *cfs_rq;
struct sched_entity *se = &p->se,
*topse = NULL; /* Highest schedulable entity */
int incload = 1;
for_each_sched_entity(se) {
topse = se;
if (se->on_rq) {
incload = 0;
break;
}
cfs_rq = cfs_rq_of(se);
enqueue_entity(cfs_rq, se, wakeup);
wakeup = 1;
}
/* Increment cpu load if we just enqueued the first task of a group on
* 'rq->cpu'. 'topse' represents the group to which task 'p' belongs
* at the highest grouping level.
*/
if (incload)
inc_cpu_load(rq, topse->load.weight);
hrtick_start_fair(rq, rq->curr);
}
/*
* The dequeue_task method is called before nr_running is
* decreased. We remove the task from the rbtree and
* update the fair scheduling stats:
*/
static void dequeue_task_fair(struct rq *rq, struct task_struct *p, int sleep)
{
struct cfs_rq *cfs_rq;
struct sched_entity *se = &p->se,
*topse = NULL; /* Highest schedulable entity */
int decload = 1;
for_each_sched_entity(se) {
topse = se;
cfs_rq = cfs_rq_of(se);
dequeue_entity(cfs_rq, se, sleep);
/* Don't dequeue parent if it has other entities besides us */
if (cfs_rq->load.weight) {
if (parent_entity(se))
decload = 0;
break;
}
sleep = 1;
}
/* Decrement cpu load if we just dequeued the last task of a group on
* 'rq->cpu'. 'topse' represents the group to which task 'p' belongs
* at the highest grouping level.
*/
if (decload)
dec_cpu_load(rq, topse->load.weight);
hrtick_start_fair(rq, rq->curr);
}
/*
* sched_yield() support is very simple - we dequeue and enqueue.
*
* If compat_yield is turned on then we requeue to the end of the tree.
*/
static void yield_task_fair(struct rq *rq)
{
struct task_struct *curr = rq->curr;
struct cfs_rq *cfs_rq = task_cfs_rq(curr);
struct sched_entity *rightmost, *se = &curr->se;
/*
* Are we the only task in the tree?
*/
if (unlikely(cfs_rq->nr_running == 1))
return;
if (likely(!sysctl_sched_compat_yield) && curr->policy != SCHED_BATCH) {
__update_rq_clock(rq);
/*
* Update run-time statistics of the 'current'.
*/
update_curr(cfs_rq);
return;
}
/*
* Find the rightmost entry in the rbtree:
*/
rightmost = __pick_last_entity(cfs_rq);
/*
* Already in the rightmost position?
*/
if (unlikely(rightmost->vruntime < se->vruntime))
return;
/*
* Minimally necessary key value to be last in the tree:
* Upon rescheduling, sched_class::put_prev_task() will place
* 'current' within the tree based on its new key value.
*/
se->vruntime = rightmost->vruntime + 1;
}
/*
* wake_idle() will wake a task on an idle cpu if task->cpu is
* not idle and an idle cpu is available. The span of cpus to
* search starts with cpus closest then further out as needed,
* so we always favor a closer, idle cpu.
*
* Returns the CPU we should wake onto.
*/
#if defined(ARCH_HAS_SCHED_WAKE_IDLE)
static int wake_idle(int cpu, struct task_struct *p)
{
cpumask_t tmp;
struct sched_domain *sd;
int i;
/*
* If it is idle, then it is the best cpu to run this task.
*
* This cpu is also the best, if it has more than one task already.
* Siblings must be also busy(in most cases) as they didn't already
* pickup the extra load from this cpu and hence we need not check
* sibling runqueue info. This will avoid the checks and cache miss
* penalities associated with that.
*/
if (idle_cpu(cpu) || cpu_rq(cpu)->nr_running > 1)
return cpu;
for_each_domain(cpu, sd) {
if (sd->flags & SD_WAKE_IDLE) {
cpus_and(tmp, sd->span, p->cpus_allowed);
for_each_cpu_mask(i, tmp) {
if (idle_cpu(i)) {
if (i != task_cpu(p)) {
schedstat_inc(p,
se.nr_wakeups_idle);
}
return i;
}
}
} else {
break;
}
}
return cpu;
}
#else
static inline int wake_idle(int cpu, struct task_struct *p)
{
return cpu;
}
#endif
#ifdef CONFIG_SMP
static int select_task_rq_fair(struct task_struct *p, int sync)
{
int cpu, this_cpu;
struct rq *rq;
struct sched_domain *sd, *this_sd = NULL;
int new_cpu;
cpu = task_cpu(p);
rq = task_rq(p);
this_cpu = smp_processor_id();
new_cpu = cpu;
if (cpu == this_cpu)
goto out_set_cpu;
for_each_domain(this_cpu, sd) {
if (cpu_isset(cpu, sd->span)) {
this_sd = sd;
break;
}
}
if (unlikely(!cpu_isset(this_cpu, p->cpus_allowed)))
goto out_set_cpu;
/*
* Check for affine wakeup and passive balancing possibilities.
*/
if (this_sd) {
int idx = this_sd->wake_idx;
unsigned int imbalance;
unsigned long load, this_load;
imbalance = 100 + (this_sd->imbalance_pct - 100) / 2;
load = source_load(cpu, idx);
this_load = target_load(this_cpu, idx);
new_cpu = this_cpu; /* Wake to this CPU if we can */
if (this_sd->flags & SD_WAKE_AFFINE) {
unsigned long tl = this_load;
unsigned long tl_per_task;
/*
* Attract cache-cold tasks on sync wakeups:
*/
if (sync && !task_hot(p, rq->clock, this_sd))
goto out_set_cpu;
schedstat_inc(p, se.nr_wakeups_affine_attempts);
tl_per_task = cpu_avg_load_per_task(this_cpu);
/*
* If sync wakeup then subtract the (maximum possible)
* effect of the currently running task from the load
* of the current CPU:
*/
if (sync)
tl -= current->se.load.weight;
if ((tl <= load &&
tl + target_load(cpu, idx) <= tl_per_task) ||
100*(tl + p->se.load.weight) <= imbalance*load) {
/*
* This domain has SD_WAKE_AFFINE and
* p is cache cold in this domain, and
* there is no bad imbalance.
*/
schedstat_inc(this_sd, ttwu_move_affine);
schedstat_inc(p, se.nr_wakeups_affine);
goto out_set_cpu;
}
}
/*
* Start passive balancing when half the imbalance_pct
* limit is reached.
*/
if (this_sd->flags & SD_WAKE_BALANCE) {
if (imbalance*this_load <= 100*load) {
schedstat_inc(this_sd, ttwu_move_balance);
schedstat_inc(p, se.nr_wakeups_passive);
goto out_set_cpu;
}
}
}
new_cpu = cpu; /* Could not wake to this_cpu. Wake to cpu instead */
out_set_cpu:
return wake_idle(new_cpu, p);
}
#endif /* CONFIG_SMP */
/*
* Preempt the current task with a newly woken task if needed:
*/
static void check_preempt_wakeup(struct rq *rq, struct task_struct *p)
{
struct task_struct *curr = rq->curr;
struct cfs_rq *cfs_rq = task_cfs_rq(curr);
struct sched_entity *se = &curr->se, *pse = &p->se;
unsigned long gran;
if (unlikely(rt_prio(p->prio))) {
update_rq_clock(rq);
update_curr(cfs_rq);
resched_task(curr);
return;
}
/*
* Batch tasks do not preempt (their preemption is driven by
* the tick):
*/
if (unlikely(p->policy == SCHED_BATCH))
return;
if (!sched_feat(WAKEUP_PREEMPT))
return;
while (!is_same_group(se, pse)) {
se = parent_entity(se);
pse = parent_entity(pse);
}
gran = sysctl_sched_wakeup_granularity;
if (unlikely(se->load.weight != NICE_0_LOAD))
gran = calc_delta_fair(gran, &se->load);
if (pse->vruntime + gran < se->vruntime)
resched_task(curr);
}
static struct task_struct *pick_next_task_fair(struct rq *rq)
{
struct task_struct *p;
struct cfs_rq *cfs_rq = &rq->cfs;
struct sched_entity *se;
if (unlikely(!cfs_rq->nr_running))
return NULL;
do {
se = pick_next_entity(cfs_rq);
cfs_rq = group_cfs_rq(se);
} while (cfs_rq);
p = task_of(se);
hrtick_start_fair(rq, p);
return p;
}
/*
* Account for a descheduled task:
*/
static void put_prev_task_fair(struct rq *rq, struct task_struct *prev)
{
struct sched_entity *se = &prev->se;
struct cfs_rq *cfs_rq;
for_each_sched_entity(se) {
cfs_rq = cfs_rq_of(se);
put_prev_entity(cfs_rq, se);
}
}
#ifdef CONFIG_SMP
/**************************************************
* Fair scheduling class load-balancing methods:
*/
/*
* Load-balancing iterator. Note: while the runqueue stays locked
* during the whole iteration, the current task might be
* dequeued so the iterator has to be dequeue-safe. Here we
* achieve that by always pre-iterating before returning
* the current task:
*/
static struct task_struct *
__load_balance_iterator(struct cfs_rq *cfs_rq, struct rb_node *curr)
{
struct task_struct *p;
if (!curr)
return NULL;
p = rb_entry(curr, struct task_struct, se.run_node);
cfs_rq->rb_load_balance_curr = rb_next(curr);
return p;
}
static struct task_struct *load_balance_start_fair(void *arg)
{
struct cfs_rq *cfs_rq = arg;
return __load_balance_iterator(cfs_rq, first_fair(cfs_rq));
}
static struct task_struct *load_balance_next_fair(void *arg)
{
struct cfs_rq *cfs_rq = arg;
return __load_balance_iterator(cfs_rq, cfs_rq->rb_load_balance_curr);
}
sched: simplify move_tasks() The move_tasks() function is currently multiplexed with two distinct capabilities: 1. attempt to move a specified amount of weighted load from one run queue to another; and 2. attempt to move a specified number of tasks from one run queue to another. The first of these capabilities is used in two places, load_balance() and load_balance_idle(), and in both of these cases the return value of move_tasks() is used purely to decide if tasks/load were moved and no notice of the actual number of tasks moved is taken. The second capability is used in exactly one place, active_load_balance(), to attempt to move exactly one task and, as before, the return value is only used as an indicator of success or failure. This multiplexing of sched_task() was introduced, by me, as part of the smpnice patches and was motivated by the fact that the alternative, one function to move specified load and one to move a single task, would have led to two functions of roughly the same complexity as the old move_tasks() (or the new balance_tasks()). However, the new modular design of the new CFS scheduler allows a simpler solution to be adopted and this patch addresses that solution by: 1. adding a new function, move_one_task(), to be used by active_load_balance(); and 2. making move_tasks() a single purpose function that tries to move a specified weighted load and returns 1 for success and 0 for failure. One of the consequences of these changes is that neither move_one_task() or the new move_tasks() care how many tasks sched_class.load_balance() moves and this enables its interface to be simplified by returning the amount of load moved as its result and removing the load_moved pointer from the argument list. This helps simplify the new move_tasks() and slightly reduces the amount of work done in each of sched_class.load_balance()'s implementations. Further simplification, e.g. changes to balance_tasks(), are possible but (slightly) complicated by the special needs of load_balance_fair() so I've left them to a later patch (if this one gets accepted). NB Since move_tasks() gets called with two run queue locks held even small reductions in overhead are worthwhile. [ mingo@elte.hu ] this change also reduces code size nicely: text data bss dec hex filename 39216 3618 24 42858 a76a sched.o.before 39173 3618 24 42815 a73f sched.o.after Signed-off-by: Peter Williams <pwil3058@bigpond.net.au> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2007-08-09 09:16:46 +00:00
static unsigned long
load_balance_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
unsigned long max_load_move,
sched: fix bug in balance_tasks() There are two problems with balance_tasks() and how it used: 1. The variables best_prio and best_prio_seen (inherited from the old move_tasks()) were only required to handle problems caused by the active/expired arrays, the order in which they were processed and the possibility that the task with the highest priority could be on either. These issues are no longer present and the extra overhead associated with their use is unnecessary (and possibly wrong). 2. In the absence of CONFIG_FAIR_GROUP_SCHED being set, the same this_best_prio variable needs to be used by all scheduling classes or there is a risk of moving too much load. E.g. if the highest priority task on this at the beginning is a fairly low priority task and the rt class migrates a task (during its turn) then that moved task becomes the new highest priority task on this_rq but when the sched_fair class initializes its copy of this_best_prio it will get the priority of the original highest priority task as, due to the run queue locks being held, the reschedule triggered by pull_task() will not have taken place. This could result in inappropriate overriding of skip_for_load and excessive load being moved. The attached patch addresses these problems by deleting all reference to best_prio and best_prio_seen and making this_best_prio a reference parameter to the various functions involved. load_balance_fair() has also been modified so that this_best_prio is only reset (in the loop) if CONFIG_FAIR_GROUP_SCHED is set. This should preserve the effect of helping spread groups' higher priority tasks around the available CPUs while improving system performance when CONFIG_FAIR_GROUP_SCHED isn't set. Signed-off-by: Peter Williams <pwil3058@bigpond.net.au> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2007-08-09 09:16:46 +00:00
struct sched_domain *sd, enum cpu_idle_type idle,
int *all_pinned, int *this_best_prio)
{
struct cfs_rq *busy_cfs_rq;
long rem_load_move = max_load_move;
struct rq_iterator cfs_rq_iterator;
sched: group scheduler, fix fairness of cpu bandwidth allocation for task groups 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>
2008-01-25 20:08:00 +00:00
unsigned long load_moved;
cfs_rq_iterator.start = load_balance_start_fair;
cfs_rq_iterator.next = load_balance_next_fair;
for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
sched: fix bug in balance_tasks() There are two problems with balance_tasks() and how it used: 1. The variables best_prio and best_prio_seen (inherited from the old move_tasks()) were only required to handle problems caused by the active/expired arrays, the order in which they were processed and the possibility that the task with the highest priority could be on either. These issues are no longer present and the extra overhead associated with their use is unnecessary (and possibly wrong). 2. In the absence of CONFIG_FAIR_GROUP_SCHED being set, the same this_best_prio variable needs to be used by all scheduling classes or there is a risk of moving too much load. E.g. if the highest priority task on this at the beginning is a fairly low priority task and the rt class migrates a task (during its turn) then that moved task becomes the new highest priority task on this_rq but when the sched_fair class initializes its copy of this_best_prio it will get the priority of the original highest priority task as, due to the run queue locks being held, the reschedule triggered by pull_task() will not have taken place. This could result in inappropriate overriding of skip_for_load and excessive load being moved. The attached patch addresses these problems by deleting all reference to best_prio and best_prio_seen and making this_best_prio a reference parameter to the various functions involved. load_balance_fair() has also been modified so that this_best_prio is only reset (in the loop) if CONFIG_FAIR_GROUP_SCHED is set. This should preserve the effect of helping spread groups' higher priority tasks around the available CPUs while improving system performance when CONFIG_FAIR_GROUP_SCHED isn't set. Signed-off-by: Peter Williams <pwil3058@bigpond.net.au> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2007-08-09 09:16:46 +00:00
#ifdef CONFIG_FAIR_GROUP_SCHED
sched: group scheduler, fix fairness of cpu bandwidth allocation for task groups 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>
2008-01-25 20:08:00 +00:00
struct cfs_rq *this_cfs_rq = busy_cfs_rq->tg->cfs_rq[this_cpu];
unsigned long maxload, task_load, group_weight;
unsigned long thisload, per_task_load;
struct sched_entity *se = busy_cfs_rq->tg->se[busiest->cpu];
sched: group scheduler, fix fairness of cpu bandwidth allocation for task groups 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>
2008-01-25 20:08:00 +00:00
task_load = busy_cfs_rq->load.weight;
group_weight = se->load.weight;
sched: group scheduler, fix fairness of cpu bandwidth allocation for task groups 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>
2008-01-25 20:08:00 +00:00
/*
* 'group_weight' is contributed by tasks of total weight
* 'task_load'. To move 'rem_load_move' worth of weight only,
* we need to move a maximum task load of:
*
* maxload = (remload / group_weight) * task_load;
*/
maxload = (rem_load_move * task_load) / group_weight;
if (!maxload || !task_load)
continue;
sched: group scheduler, fix fairness of cpu bandwidth allocation for task groups 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>
2008-01-25 20:08:00 +00:00
per_task_load = task_load / busy_cfs_rq->nr_running;
/*
* balance_tasks will try to forcibly move atleast one task if
* possible (because of SCHED_LOAD_SCALE_FUZZ). Avoid that if
* maxload is less than GROUP_IMBALANCE_FUZZ% the per_task_load.
*/
if (100 * maxload < GROUP_IMBALANCE_PCT * per_task_load)
continue;
sched: group scheduler, fix fairness of cpu bandwidth allocation for task groups 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>
2008-01-25 20:08:00 +00:00
/* Disable priority-based load balance */
*this_best_prio = 0;
thisload = this_cfs_rq->load.weight;
sched: fix bug in balance_tasks() There are two problems with balance_tasks() and how it used: 1. The variables best_prio and best_prio_seen (inherited from the old move_tasks()) were only required to handle problems caused by the active/expired arrays, the order in which they were processed and the possibility that the task with the highest priority could be on either. These issues are no longer present and the extra overhead associated with their use is unnecessary (and possibly wrong). 2. In the absence of CONFIG_FAIR_GROUP_SCHED being set, the same this_best_prio variable needs to be used by all scheduling classes or there is a risk of moving too much load. E.g. if the highest priority task on this at the beginning is a fairly low priority task and the rt class migrates a task (during its turn) then that moved task becomes the new highest priority task on this_rq but when the sched_fair class initializes its copy of this_best_prio it will get the priority of the original highest priority task as, due to the run queue locks being held, the reschedule triggered by pull_task() will not have taken place. This could result in inappropriate overriding of skip_for_load and excessive load being moved. The attached patch addresses these problems by deleting all reference to best_prio and best_prio_seen and making this_best_prio a reference parameter to the various functions involved. load_balance_fair() has also been modified so that this_best_prio is only reset (in the loop) if CONFIG_FAIR_GROUP_SCHED is set. This should preserve the effect of helping spread groups' higher priority tasks around the available CPUs while improving system performance when CONFIG_FAIR_GROUP_SCHED isn't set. Signed-off-by: Peter Williams <pwil3058@bigpond.net.au> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2007-08-09 09:16:46 +00:00
#else
# define maxload rem_load_move
sched: fix bug in balance_tasks() There are two problems with balance_tasks() and how it used: 1. The variables best_prio and best_prio_seen (inherited from the old move_tasks()) were only required to handle problems caused by the active/expired arrays, the order in which they were processed and the possibility that the task with the highest priority could be on either. These issues are no longer present and the extra overhead associated with their use is unnecessary (and possibly wrong). 2. In the absence of CONFIG_FAIR_GROUP_SCHED being set, the same this_best_prio variable needs to be used by all scheduling classes or there is a risk of moving too much load. E.g. if the highest priority task on this at the beginning is a fairly low priority task and the rt class migrates a task (during its turn) then that moved task becomes the new highest priority task on this_rq but when the sched_fair class initializes its copy of this_best_prio it will get the priority of the original highest priority task as, due to the run queue locks being held, the reschedule triggered by pull_task() will not have taken place. This could result in inappropriate overriding of skip_for_load and excessive load being moved. The attached patch addresses these problems by deleting all reference to best_prio and best_prio_seen and making this_best_prio a reference parameter to the various functions involved. load_balance_fair() has also been modified so that this_best_prio is only reset (in the loop) if CONFIG_FAIR_GROUP_SCHED is set. This should preserve the effect of helping spread groups' higher priority tasks around the available CPUs while improving system performance when CONFIG_FAIR_GROUP_SCHED isn't set. Signed-off-by: Peter Williams <pwil3058@bigpond.net.au> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2007-08-09 09:16:46 +00:00
#endif
/*
* pass busy_cfs_rq argument into
* load_balance_[start|next]_fair iterators
*/
cfs_rq_iterator.arg = busy_cfs_rq;
sched: group scheduler, fix fairness of cpu bandwidth allocation for task groups 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>
2008-01-25 20:08:00 +00:00
load_moved = balance_tasks(this_rq, this_cpu, busiest,
maxload, sd, idle, all_pinned,
this_best_prio,
&cfs_rq_iterator);
sched: group scheduler, fix fairness of cpu bandwidth allocation for task groups 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>
2008-01-25 20:08:00 +00:00
#ifdef CONFIG_FAIR_GROUP_SCHED
/*
* load_moved holds the task load that was moved. The
* effective (group) weight moved would be:
* load_moved_eff = load_moved/task_load * group_weight;
*/
load_moved = (group_weight * load_moved) / task_load;
/* Adjust shares on both cpus to reflect load_moved */
group_weight -= load_moved;
set_se_shares(se, group_weight);
se = busy_cfs_rq->tg->se[this_cpu];
if (!thisload)
group_weight = load_moved;
else
group_weight = se->load.weight + load_moved;
set_se_shares(se, group_weight);
#endif
rem_load_move -= load_moved;
if (rem_load_move <= 0)
break;
}
sched: simplify move_tasks() The move_tasks() function is currently multiplexed with two distinct capabilities: 1. attempt to move a specified amount of weighted load from one run queue to another; and 2. attempt to move a specified number of tasks from one run queue to another. The first of these capabilities is used in two places, load_balance() and load_balance_idle(), and in both of these cases the return value of move_tasks() is used purely to decide if tasks/load were moved and no notice of the actual number of tasks moved is taken. The second capability is used in exactly one place, active_load_balance(), to attempt to move exactly one task and, as before, the return value is only used as an indicator of success or failure. This multiplexing of sched_task() was introduced, by me, as part of the smpnice patches and was motivated by the fact that the alternative, one function to move specified load and one to move a single task, would have led to two functions of roughly the same complexity as the old move_tasks() (or the new balance_tasks()). However, the new modular design of the new CFS scheduler allows a simpler solution to be adopted and this patch addresses that solution by: 1. adding a new function, move_one_task(), to be used by active_load_balance(); and 2. making move_tasks() a single purpose function that tries to move a specified weighted load and returns 1 for success and 0 for failure. One of the consequences of these changes is that neither move_one_task() or the new move_tasks() care how many tasks sched_class.load_balance() moves and this enables its interface to be simplified by returning the amount of load moved as its result and removing the load_moved pointer from the argument list. This helps simplify the new move_tasks() and slightly reduces the amount of work done in each of sched_class.load_balance()'s implementations. Further simplification, e.g. changes to balance_tasks(), are possible but (slightly) complicated by the special needs of load_balance_fair() so I've left them to a later patch (if this one gets accepted). NB Since move_tasks() gets called with two run queue locks held even small reductions in overhead are worthwhile. [ mingo@elte.hu ] this change also reduces code size nicely: text data bss dec hex filename 39216 3618 24 42858 a76a sched.o.before 39173 3618 24 42815 a73f sched.o.after Signed-off-by: Peter Williams <pwil3058@bigpond.net.au> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2007-08-09 09:16:46 +00:00
return max_load_move - rem_load_move;
}
static int
move_one_task_fair(struct rq *this_rq, int this_cpu, struct rq *busiest,
struct sched_domain *sd, enum cpu_idle_type idle)
{
struct cfs_rq *busy_cfs_rq;
struct rq_iterator cfs_rq_iterator;
cfs_rq_iterator.start = load_balance_start_fair;
cfs_rq_iterator.next = load_balance_next_fair;
for_each_leaf_cfs_rq(busiest, busy_cfs_rq) {
/*
* pass busy_cfs_rq argument into
* load_balance_[start|next]_fair iterators
*/
cfs_rq_iterator.arg = busy_cfs_rq;
if (iter_move_one_task(this_rq, this_cpu, busiest, sd, idle,
&cfs_rq_iterator))
return 1;
}
return 0;
}
#endif
/*
* scheduler tick hitting a task of our scheduling class:
*/
static void task_tick_fair(struct rq *rq, struct task_struct *curr, int queued)
{
struct cfs_rq *cfs_rq;
struct sched_entity *se = &curr->se;
for_each_sched_entity(se) {
cfs_rq = cfs_rq_of(se);
entity_tick(cfs_rq, se, queued);
}
}
#define swap(a, b) do { typeof(a) tmp = (a); (a) = (b); (b) = tmp; } while (0)
/*
* Share the fairness runtime between parent and child, thus the
* total amount of pressure for CPU stays equal - new tasks
* get a chance to run but frequent forkers are not allowed to
* monopolize the CPU. Note: the parent runqueue is locked,
* the child is not running yet.
*/
static void task_new_fair(struct rq *rq, struct task_struct *p)
{
struct cfs_rq *cfs_rq = task_cfs_rq(p);
struct sched_entity *se = &p->se, *curr = cfs_rq->curr;
int this_cpu = smp_processor_id();
sched_info_queued(p);
update_curr(cfs_rq);
place_entity(cfs_rq, se, 1);
sched: fix copy_namespace() <-> sched_fork() dependency in do_fork 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>
2007-11-09 21:39:39 +00:00
/* 'curr' will be NULL if the child belongs to a different group */
if (sysctl_sched_child_runs_first && this_cpu == task_cpu(p) &&
sched: fix copy_namespace() <-> sched_fork() dependency in do_fork 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>
2007-11-09 21:39:39 +00:00
curr && curr->vruntime < se->vruntime) {
/*
* Upon rescheduling, sched_class::put_prev_task() will place
* 'current' within the tree based on its new key value.
*/
swap(curr->vruntime, se->vruntime);
}
enqueue_task_fair(rq, p, 0);
resched_task(rq->curr);
}
/*
* Priority of the task has changed. Check to see if we preempt
* the current task.
*/
static void prio_changed_fair(struct rq *rq, struct task_struct *p,
int oldprio, int running)
{
/*
* Reschedule if we are currently running on this runqueue and
* our priority decreased, or if we are not currently running on
* this runqueue and our priority is higher than the current's
*/
if (running) {
if (p->prio > oldprio)
resched_task(rq->curr);
} else
check_preempt_curr(rq, p);
}
/*
* We switched to the sched_fair class.
*/
static void switched_to_fair(struct rq *rq, struct task_struct *p,
int running)
{
/*
* We were most likely switched from sched_rt, so
* kick off the schedule if running, otherwise just see
* if we can still preempt the current task.
*/
if (running)
resched_task(rq->curr);
else
check_preempt_curr(rq, p);
}
/* Account for a task changing its policy or group.
*
* This routine is mostly called to set cfs_rq->curr field when a task
* migrates between groups/classes.
*/
static void set_curr_task_fair(struct rq *rq)
{
struct sched_entity *se = &rq->curr->se;
for_each_sched_entity(se)
set_next_entity(cfs_rq_of(se), se);
}
/*
* All the scheduling class methods:
*/
static const struct sched_class fair_sched_class = {
.next = &idle_sched_class,
.enqueue_task = enqueue_task_fair,
.dequeue_task = dequeue_task_fair,
.yield_task = yield_task_fair,
#ifdef CONFIG_SMP
.select_task_rq = select_task_rq_fair,
#endif /* CONFIG_SMP */
.check_preempt_curr = check_preempt_wakeup,
.pick_next_task = pick_next_task_fair,
.put_prev_task = put_prev_task_fair,
#ifdef CONFIG_SMP
.load_balance = load_balance_fair,
.move_one_task = move_one_task_fair,
#endif
.set_curr_task = set_curr_task_fair,
.task_tick = task_tick_fair,
.task_new = task_new_fair,
.prio_changed = prio_changed_fair,
.switched_to = switched_to_fair,
};
#ifdef CONFIG_SCHED_DEBUG
static void print_cfs_stats(struct seq_file *m, int cpu)
{
struct cfs_rq *cfs_rq;
#ifdef CONFIG_FAIR_GROUP_SCHED
print_cfs_rq(m, cpu, &cpu_rq(cpu)->cfs);
#endif
rcu_read_lock();
for_each_leaf_cfs_rq(cpu_rq(cpu), cfs_rq)
print_cfs_rq(m, cpu, cfs_rq);
rcu_read_unlock();
}
#endif