lkml.org 
[lkml]   [2011]   [Aug]   [2]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
Subject[PATCH][GIT PULL] sched/cpupri: Remove the vec->lock
From
Date

Ingo,

I've been passing this patch (one with benchmark code enabled) around,
and this has shown nice improvement in the scheduling of RT tasks. The
original code in cpupri, uses a vec->lock where there's one vec per RT
priority, and these locks are global. With the RCU threads and IRQ
threads being RT tasks of the same priority, this lock has been showing
up at the top of the contention list. It has been causing some severe
performance hits in some cases on large boxes.

I passed this patch around to various users that have access to
different boxes ranging from 4 to 64 CPUs. In every case, this patch
showed a significant improvement, as it replaces the global vec->lock
with memory barriers. The added measurement tests is not the only
improvements, but so are jitter tests and cyclictest have improved when
this patch (without the benchmark recording) has been applied.
This patch is in my RT git repo based off of v3.0.


Please pull the latest tip/sched/cpupri tree, which can be found at:

git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-2.6-rt.git
tip/sched/cpupri

Head SHA1: 84d342f89dd0bf7d9a01c9802021f01a5ff1c453


Steven Rostedt (1):
sched/cpupri: Remove the vec->lock

----
kernel/sched_cpupri.c | 62 ++++++++++++++++++++++++++++++------------------
kernel/sched_cpupri.h | 5 +--
2 files changed, 41 insertions(+), 26 deletions(-)
---------------------------
commit 84d342f89dd0bf7d9a01c9802021f01a5ff1c453
Author: Steven Rostedt <srostedt@redhat.com>
Date: Mon Aug 1 13:20:05 2011 -0400
sched/cpupri: Remove the vec->lock

The cpupri vec->lock has been showing up as a top contention
lately. This is because of the RT push/pull logic takes an
agressive approach for migrating RT tasks. The cpupri logic is
in place to improve the performance of the push/pull when dealing
with large number CPU machines.

The problem though is a vec->lock is required, where a vec is a
global per RT priority structure. That is, if there are lots of
RT tasks at the same priority, every time they are added or removed
from the RT queue, this global vec->lock is taken. Now that more
kernel threads are becoming RT (RCU boost and threaded interrupts)
this is becoming much more of an issue.

There are two variables that are being synced by the vec->lock.
The cpupri bitmask, and the vec->counter. The cpupri bitmask
is one bit per priority. If a RT priority vec has a process queued,
then the vec->count is > 0 and the cpupri bitmask is set for that
RT priority.
If the cpupri bitmask gets out of sync with the vec->counter, we could
end up pushing a low proirity RT task to a high priority queue.
That RT task that could have run immediately could be queued on a
run queue with a higher priority task indefinitely.

The solution is not to use the cpupri bitmask and just look at the
vec->count directly when doing a pull. The cpupri bitmask is just
a fast way to scan the RT priorities when a pull is made. Instead
of using the bitmask, and just examine all RT priorities, and
look at the vec->counts, we could eliminate the vec->lock. The
scan of RT tasks is to find a run queue that we can push an RT task
to, and we do not push to a high priority queue, thus the scan only
needs to go from 1 to RT task->prio, and not all 100 RT priorities.

The push algorithm, which does the scan of RT priorities (and
scan of the bitmask) only happens when we have an overloaded RT run
queue (more than one RT task queued). The grabbing of the vec->lock
happens every time any RT task is queued or dequeued on the run
queue for that priority. The slowing down of the scan by not using
a bitmask is negligible by the speed up of removing the vec->lock
contention, and replacing it with an atomic counter and memory barrier.

To prove this, I wrote a patch that times both the loop and the code
that grabs the vec->locks. I passed the patches to various people
(and companies) to test and show the results. I let everyone choose
their own load to test, giving different loads on the system,
for various different setups.

Here's some of the results: (snipping to a few CPUs to not make
this change log huge, but the results were consistent across
the entire system).

System 1 (24 CPUs)

Before patch:
CPU: Name Count Max Min Average Total
---- ---- ----- --- --- ------- -----
[...]
cpu 20: loop 3057 1.766 0.061 0.642 1963.170
vec 6782949 90.469 0.089 0.414 2811760.503
cpu 21: loop 2617 1.723 0.062 0.641 1679.074
vec 6782810 90.499 0.089 0.291 1978499.900
cpu 22: loop 2212 1.863 0.063 0.699 1547.160
vec 6767244 85.685 0.089 0.435 2949676.898
cpu 23: loop 2320 2.013 0.062 0.594 1380.265
vec 6781694 87.923 0.088 0.431 2928538.224
After patch:
cpu 20: loop 2078 1.579 0.061 0.533 1108.006
vec 6164555 5.704 0.060 0.143 885185.809
cpu 21: loop 2268 1.712 0.065 0.575 1305.248
vec 6153376 5.558 0.060 0.187 1154960.469
cpu 22: loop 1542 1.639 0.095 0.533 823.249
vec 6156510 5.720 0.060 0.190 1172727.232
cpu 23: loop 1650 1.733 0.068 0.545 900.781
vec 6170784 5.533 0.060 0.167 1034287.953

All times are in microseconds. The 'loop' is the amount of time spent
doing the loop across the priorities (before patch uses bitmask).
the 'vec' is the amount of time in the code that requires grabbing
the vec->lock. The second patch just does not have the vec lock, but
encompasses the same code.

Amazingly the loop code even went down on average. The vec code went
from .5 down to .18, that's more than half the time spent!

Note, more than one test was run, but they all had the same results.

System 2 (64 CPUs)

Before patch:
CPU: Name Count Max Min Average Total
---- ---- ----- --- --- ------- -----
cpu 60: loop 0 0 0 0 0
vec 5410840 277.954 0.084 0.782 4232895.727
cpu 61: loop 0 0 0 0 0
vec 4915648 188.399 0.084 0.570 2803220.301
cpu 62: loop 0 0 0 0 0
vec 5356076 276.417 0.085 0.786 4214544.548
cpu 63: loop 0 0 0 0 0
vec 4891837 170.531 0.085 0.799 3910948.833
After patch:
cpu 60: loop 0 0 0 0 0
vec 5365118 5.080 0.021 0.063 340490.267
cpu 61: loop 0 0 0 0 0
vec 4898590 1.757 0.019 0.071 347903.615
cpu 62: loop 0 0 0 0 0
vec 5737130 3.067 0.021 0.119 687108.734
cpu 63: loop 0 0 0 0 0
vec 4903228 1.822 0.021 0.071 348506.477

The test run during the measurement did not have any (very few,
from other CPUs) RT tasks pushing. But this shows that it helped
out tremendously with the contention, as the contention happens
because the vec->lock is taken only on queuing at an RT priority,
and different CPUs that queue tasks at the same priority will
have contention.

I tested on my own 4 CPU machine with the following results:

Before patch:
CPU: Name Count Max Min Average Total
---- ---- ----- --- --- ------- -----
cpu 0: loop 2377 1.489 0.158 0.588 1398.395
vec 4484 770.146 2.301 4.396 19711.755
cpu 1: loop 2169 1.962 0.160 0.576 1250.110
vec 4425 152.769 2.297 4.030 17834.228
cpu 2: loop 2324 1.749 0.155 0.559 1299.799
vec 4368 779.632 2.325 4.665 20379.268
cpu 3: loop 2325 1.629 0.157 0.561 1306.113
vec 4650 408.782 2.394 4.348 20222.577
After patch:
CPU: Name Count Max Min Average Total
---- ---- ----- --- --- ------- -----
cpu 0: loop 2121 1.616 0.113 0.636 1349.189
vec 4303 1.151 0.225 0.421 1811.966
cpu 1: loop 2130 1.638 0.178 0.644 1372.927
vec 4627 1.379 0.235 0.428 1983.648
cpu 2: loop 2056 1.464 0.165 0.637 1310.141
vec 4471 1.311 0.217 0.433 1937.927
cpu 3: loop 2154 1.481 0.162 0.601 1295.083
vec 4236 1.253 0.230 0.425 1803.008
This was running my migrate.c code that can be found at:
http://lwn.net/Articles/425763/

The migrate code does stress the RT tasks a bit. This shows that
the loop did increase a little after the patch, but not by much.
The vec code dropped dramatically. From 4.3us down to .42us.
That's a 10x improvement!

Tested-by: Mike Galbraith <mgalbraith@suse.de>
Tested-by: Luis Claudio R. Gonçalves <lgoncalv@redhat.com>
Tested-by: Matthew Hank Sabins<msabins@linux.vnet.ibm.com>
Reviewed-by: Gregory Haskins <gregory.haskins@gmail.com>
Signed-off-by: Steven Rostedt <rostedt@goodmis.org>

diff --git a/kernel/sched_cpupri.c b/kernel/sched_cpupri.c
index 2722dc1..7761a26 100644
--- a/kernel/sched_cpupri.c
+++ b/kernel/sched_cpupri.c
@@ -47,9 +47,6 @@ static int convert_prio(int prio)
return cpupri;
}

-#define for_each_cpupri_active(array, idx) \
- for_each_set_bit(idx, array, CPUPRI_NR_PRIORITIES)
-
/**
* cpupri_find - find the best (lowest-pri) CPU in the system
* @cp: The cpupri context
@@ -71,11 +68,33 @@ int cpupri_find(struct cpupri *cp, struct task_struct *p,
int idx = 0;
int task_pri = convert_prio(p->prio);

- for_each_cpupri_active(cp->pri_active, idx) {
+ if (task_pri >= MAX_RT_PRIO)
+ return 0;
+
+ for (idx = 0; idx < task_pri; idx++) {
struct cpupri_vec *vec = &cp->pri_to_cpu[idx];

- if (idx >= task_pri)
- break;
+ if (!atomic_read(&(vec)->count))
+ continue;
+ /*
+ * When looking at the vector, we need to read the counter,
+ * do a memory barrier, then read the mask.
+ *
+ * Note: This is still all racey, but we can deal with it.
+ * Ideally, we only want to look at masks that are set.
+ *
+ * If a mask is not set, then the only thing wrong is that we
+ * did a little more work than necessary.
+ *
+ * If we read a zero count but the mask is set, because of the
+ * memory barriers, that can only happen when the highest prio
+ * task for a run queue has left the run queue, in which case,
+ * it will be followed by a pull. If the task we are processing
+ * fails to find a proper place to go, that pull request will
+ * pull this task if the run queue is running at a lower
+ * priority.
+ */
+ smp_rmb();

if (cpumask_any_and(&p->cpus_allowed, vec->mask) >= nr_cpu_ids)
continue;
@@ -115,7 +134,6 @@ void cpupri_set(struct cpupri *cp, int cpu, int newpri)
{
int *currpri = &cp->cpu_to_pri[cpu];
int oldpri = *currpri;
- unsigned long flags;

newpri = convert_prio(newpri);

@@ -134,26 +152,25 @@ void cpupri_set(struct cpupri *cp, int cpu, int newpri)
if (likely(newpri != CPUPRI_INVALID)) {
struct cpupri_vec *vec = &cp->pri_to_cpu[newpri];

- raw_spin_lock_irqsave(&vec->lock, flags);
-
cpumask_set_cpu(cpu, vec->mask);
- vec->count++;
- if (vec->count == 1)
- set_bit(newpri, cp->pri_active);
-
- raw_spin_unlock_irqrestore(&vec->lock, flags);
+ /*
+ * When adding a new vector, we update the mask first,
+ * do a write memory barrier, and then update the count, to
+ * make sure the vector is visible when count is set.
+ */
+ smp_wmb();
+ atomic_inc(&(vec)->count);
}
if (likely(oldpri != CPUPRI_INVALID)) {
struct cpupri_vec *vec = &cp->pri_to_cpu[oldpri];

- raw_spin_lock_irqsave(&vec->lock, flags);
-
- vec->count--;
- if (!vec->count)
- clear_bit(oldpri, cp->pri_active);
+ /*
+ * When removing from the vector, we decrement the counter first
+ * do a memory barrier and then clear the mask.
+ */
+ atomic_dec(&(vec)->count);
+ smp_wmb();
cpumask_clear_cpu(cpu, vec->mask);
-
- raw_spin_unlock_irqrestore(&vec->lock, flags);
}

*currpri = newpri;
@@ -175,8 +192,7 @@ int cpupri_init(struct cpupri *cp)
for (i = 0; i < CPUPRI_NR_PRIORITIES; i++) {
struct cpupri_vec *vec = &cp->pri_to_cpu[i];

- raw_spin_lock_init(&vec->lock);
- vec->count = 0;
+ atomic_set(&vec->count, 0);
if (!zalloc_cpumask_var(&vec->mask, GFP_KERNEL))
goto cleanup;
}
diff --git a/kernel/sched_cpupri.h b/kernel/sched_cpupri.h
index 9fc7d38..6b4cd17 100644
--- a/kernel/sched_cpupri.h
+++ b/kernel/sched_cpupri.h
@@ -12,9 +12,8 @@
/* values 2-101 are RT priorities 0-99 */

struct cpupri_vec {
- raw_spinlock_t lock;
- int count;
- cpumask_var_t mask;
+ atomic_t count;
+ cpumask_var_t mask;
};

struct cpupri {

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2011-08-02 22:39    [from the cache]
©2003-2011 Jasper Spaans