rpms/kernel/F-11 linux-2.6.30-cpuidle-faster-io.patch, NONE, 1.1 kernel.spec, 1.1739, 1.1740
Chuck Ebbert
cebbert at fedoraproject.org
Thu Sep 24 21:28:02 UTC 2009
Author: cebbert
Update of /cvs/pkgs/rpms/kernel/F-11
In directory cvs1.fedora.phx.redhat.com:/tmp/cvs-serv14273
Modified Files:
kernel.spec
Added Files:
linux-2.6.30-cpuidle-faster-io.patch
Log Message:
Backport the cpuidle-faster-io fix from Fedora 12 to fix I/O
performance problems when reading/writing multiple disks.
linux-2.6.30-cpuidle-faster-io.patch:
drivers/cpuidle/governors/menu.c | 251 ++++++++++++++++++++++++++++++++-------
include/linux/sched.h | 4
kernel/sched.c | 13 ++
3 files changed, 229 insertions(+), 39 deletions(-)
--- NEW FILE linux-2.6.30-cpuidle-faster-io.patch ---
From: Arjan van de Ven <arjan at infradead.org>
Date: Tue, 22 Sep 2009 00:04:08 +0000 (-0700)
Subject: cpuidle: fix the menu governor to boost IO performance
X-Git-Url: http://git.kernel.org/?p=linux%2Fkernel%2Fgit%2Ftorvalds%2Flinux-2.6.git;a=commitdiff_plain;h=69d25870f20c4b2563304f2b79c5300dd60a067e
cpuidle: fix the menu governor to boost IO performance
[ backported to 2.6.30 ]
Fix the menu idle governor which balances power savings, energy efficiency
and performance impact.
The reason for a reworked governor is that there have been serious
performance issues reported with the existing code on Nehalem server
systems.
To show this I'm sure Andrew wants to see benchmark results:
(benchmark is "fio", "no cstates" is using "idle=poll")
no cstates current linux new algorithm
1 disk 107 Mb/s 85 Mb/s 105 Mb/s
2 disks 215 Mb/s 123 Mb/s 209 Mb/s
12 disks 590 Mb/s 320 Mb/s 585 Mb/s
In various power benchmark measurements, no degredation was found by our
measurement&diagnostics team. Obviously a small percentage more power was
used in the "fio" benchmark, due to the much higher performance.
While it would be a novel idea to describe the new algorithm in this
commit message, I cheaped out and described it in comments in the code
instead.
[changes since first post: spelling fixes from akpm, review feedback,
folded menu-tng into menu.c]
Signed-off-by: Arjan van de Ven <arjan at linux.intel.com>
Cc: Venkatesh Pallipadi <venkatesh.pallipadi at intel.com>
Cc: Len Brown <lenb at kernel.org>
Cc: Ingo Molnar <mingo at elte.hu>
Cc: Peter Zijlstra <a.p.zijlstra at chello.nl>
Cc: Yanmin Zhang <yanmin_zhang at linux.intel.com>
Acked-by: Ingo Molnar <mingo at elte.hu>
Signed-off-by: Andrew Morton <akpm at linux-foundation.org>
Signed-off-by: Andrew Morton <akpm at linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds at linux-foundation.org>
---
diff --git a/drivers/cpuidle/governors/menu.c b/drivers/cpuidle/governors/menu.c
index f1df59f..9f3d775 100644
--- a/drivers/cpuidle/governors/menu.c
+++ b/drivers/cpuidle/governors/menu.c
@@ -2,8 +2,12 @@
* menu.c - the menu idle governor
*
* Copyright (C) 2006-2007 Adam Belay <abelay at novell.com>
+ * Copyright (C) 2009 Intel Corporation
+ * Author:
+ * Arjan van de Ven <arjan at linux.intel.com>
*
- * This code is licenced under the GPL.
+ * This code is licenced under the GPL version 2 as described
+ * in the COPYING file that acompanies the Linux Kernel.
*/
#include <linux/kernel.h>
@@ -13,20 +17,153 @@
#include <linux/ktime.h>
#include <linux/hrtimer.h>
#include <linux/tick.h>
+#include <linux/sched.h>
-#define BREAK_FUZZ 4 /* 4 us */
-#define PRED_HISTORY_PCT 50
+#define BUCKETS 12
+#define RESOLUTION 1024
+#define DECAY 4
+#define MAX_INTERESTING 50000
+
+/*
+ * Concepts and ideas behind the menu governor
+ *
+ * For the menu governor, there are 3 decision factors for picking a C
+ * state:
+ * 1) Energy break even point
+ * 2) Performance impact
+ * 3) Latency tolerance (from pmqos infrastructure)
+ * These these three factors are treated independently.
+ *
+ * Energy break even point
+ * -----------------------
+ * C state entry and exit have an energy cost, and a certain amount of time in
+ * the C state is required to actually break even on this cost. CPUIDLE
+ * provides us this duration in the "target_residency" field. So all that we
+ * need is a good prediction of how long we'll be idle. Like the traditional
+ * menu governor, we start with the actual known "next timer event" time.
+ *
+ * Since there are other source of wakeups (interrupts for example) than
+ * the next timer event, this estimation is rather optimistic. To get a
+ * more realistic estimate, a correction factor is applied to the estimate,
+ * that is based on historic behavior. For example, if in the past the actual
+ * duration always was 50% of the next timer tick, the correction factor will
+ * be 0.5.
+ *
+ * menu uses a running average for this correction factor, however it uses a
+ * set of factors, not just a single factor. This stems from the realization
+ * that the ratio is dependent on the order of magnitude of the expected
+ * duration; if we expect 500 milliseconds of idle time the likelihood of
+ * getting an interrupt very early is much higher than if we expect 50 micro
+ * seconds of idle time. A second independent factor that has big impact on
+ * the actual factor is if there is (disk) IO outstanding or not.
+ * (as a special twist, we consider every sleep longer than 50 milliseconds
+ * as perfect; there are no power gains for sleeping longer than this)
+ *
+ * For these two reasons we keep an array of 12 independent factors, that gets
+ * indexed based on the magnitude of the expected duration as well as the
+ * "is IO outstanding" property.
+ *
+ * Limiting Performance Impact
+ * ---------------------------
+ * C states, especially those with large exit latencies, can have a real
+ * noticable impact on workloads, which is not acceptable for most sysadmins,
+ * and in addition, less performance has a power price of its own.
+ *
+ * As a general rule of thumb, menu assumes that the following heuristic
+ * holds:
+ * The busier the system, the less impact of C states is acceptable
+ *
+ * This rule-of-thumb is implemented using a performance-multiplier:
+ * If the exit latency times the performance multiplier is longer than
+ * the predicted duration, the C state is not considered a candidate
+ * for selection due to a too high performance impact. So the higher
+ * this multiplier is, the longer we need to be idle to pick a deep C
+ * state, and thus the less likely a busy CPU will hit such a deep
+ * C state.
+ *
+ * Two factors are used in determing this multiplier:
+ * a value of 10 is added for each point of "per cpu load average" we have.
+ * a value of 5 points is added for each process that is waiting for
+ * IO on this CPU.
+ * (these values are experimentally determined)
+ *
+ * The load average factor gives a longer term (few seconds) input to the
+ * decision, while the iowait value gives a cpu local instantanious input.
+ * The iowait factor may look low, but realize that this is also already
+ * represented in the system load average.
+ *
+ */
struct menu_device {
int last_state_idx;
unsigned int expected_us;
- unsigned int predicted_us;
- unsigned int current_predicted_us;
- unsigned int last_measured_us;
- unsigned int elapsed_us;
+ u64 predicted_us;
+ unsigned int measured_us;
+ unsigned int exit_us;
+ unsigned int bucket;
+ u64 correction_factor[BUCKETS];
};
+
+#define LOAD_INT(x) ((x) >> FSHIFT)
+#define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100)
+
+static int get_loadavg(void)
+{
+ unsigned long this = this_cpu_load();
+
+
+ return LOAD_INT(this) * 10 + LOAD_FRAC(this) / 10;
+}
+
+static inline int which_bucket(unsigned int duration)
+{
+ int bucket = 0;
+
+ /*
+ * We keep two groups of stats; one with no
+ * IO pending, one without.
+ * This allows us to calculate
+ * E(duration)|iowait
+ */
+ if (nr_iowait_cpu())
+ bucket = BUCKETS/2;
+
+ if (duration < 10)
+ return bucket;
+ if (duration < 100)
+ return bucket + 1;
+ if (duration < 1000)
+ return bucket + 2;
+ if (duration < 10000)
+ return bucket + 3;
+ if (duration < 100000)
+ return bucket + 4;
+ return bucket + 5;
+}
+
+/*
+ * Return a multiplier for the exit latency that is intended
+ * to take performance requirements into account.
+ * The more performance critical we estimate the system
+ * to be, the higher this multiplier, and thus the higher
+ * the barrier to go to an expensive C state.
+ */
+static inline int performance_multiplier(void)
+{
+ int mult = 1;
+
+ /* for higher loadavg, we are more reluctant */
+
+ mult += 2 * get_loadavg();
+
+ /* for IO wait tasks (per cpu!) we add 5x each */
+ mult += 10 * nr_iowait_cpu();
+
+ return mult;
+}
+
static DEFINE_PER_CPU(struct menu_device, menu_devices);
/**
@@ -38,37 +175,59 @@ static int menu_select(struct cpuidle_device *dev)
struct menu_device *data = &__get_cpu_var(menu_devices);
int latency_req = pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY);
int i;
+ int multiplier;
+
+ data->last_state_idx = 0;
+ data->exit_us = 0;
/* Special case when user has set very strict latency requirement */
- if (unlikely(latency_req == 0)) {
- data->last_state_idx = 0;
+ if (unlikely(latency_req == 0))
return 0;
- }
- /* determine the expected residency time */
+ /* determine the expected residency time, round up */
data->expected_us =
- (u32) ktime_to_ns(tick_nohz_get_sleep_length()) / 1000;
+ DIV_ROUND_UP((u32)ktime_to_ns(tick_nohz_get_sleep_length()), 1000);
+
+
+ data->bucket = which_bucket(data->expected_us);
+
+ multiplier = performance_multiplier();
+
+ /*
+ * if the correction factor is 0 (eg first time init or cpu hotplug
+ * etc), we actually want to start out with a unity factor.
+ */
+ if (data->correction_factor[data->bucket] == 0)
+ data->correction_factor[data->bucket] = RESOLUTION * DECAY;
+
+ /* Make sure to round up for half microseconds */
+ data->predicted_us = DIV_ROUND_CLOSEST(
+ data->expected_us * data->correction_factor[data->bucket],
+ RESOLUTION * DECAY);
+
+ /*
+ * We want to default to C1 (hlt), not to busy polling
+ * unless the timer is happening really really soon.
+ */
+ if (data->expected_us > 5)
+ data->last_state_idx = CPUIDLE_DRIVER_STATE_START;
- /* Recalculate predicted_us based on prediction_history_pct */
- data->predicted_us *= PRED_HISTORY_PCT;
- data->predicted_us += (100 - PRED_HISTORY_PCT) *
- data->current_predicted_us;
- data->predicted_us /= 100;
/* find the deepest idle state that satisfies our constraints */
- for (i = CPUIDLE_DRIVER_STATE_START + 1; i < dev->state_count; i++) {
+ for (i = CPUIDLE_DRIVER_STATE_START; i < dev->state_count; i++) {
struct cpuidle_state *s = &dev->states[i];
- if (s->target_residency > data->expected_us)
- break;
if (s->target_residency > data->predicted_us)
break;
if (s->exit_latency > latency_req)
break;
+ if (s->exit_latency * multiplier > data->predicted_us)
+ break;
+ data->exit_us = s->exit_latency;
+ data->last_state_idx = i;
}
- data->last_state_idx = i - 1;
- return i - 1;
+ return data->last_state_idx;
}
/**
@@ -85,35 +244,49 @@ static void menu_reflect(struct cpuidle_device *dev)
unsigned int last_idle_us = cpuidle_get_last_residency(dev);
struct cpuidle_state *target = &dev->states[last_idx];
unsigned int measured_us;
+ u64 new_factor;
/*
* Ugh, this idle state doesn't support residency measurements, so we
* are basically lost in the dark. As a compromise, assume we slept
- * for one full standard timer tick. However, be aware that this
- * could potentially result in a suboptimal state transition.
+ * for the whole expected time.
*/
if (unlikely(!(target->flags & CPUIDLE_FLAG_TIME_VALID)))
- last_idle_us = USEC_PER_SEC / HZ;
+ last_idle_us = data->expected_us;
+
+
+ measured_us = last_idle_us;
/*
- * measured_us and elapsed_us are the cumulative idle time, since the
- * last time we were woken out of idle by an interrupt.
+ * We correct for the exit latency; we are assuming here that the
+ * exit latency happens after the event that we're interested in.
*/
- if (data->elapsed_us <= data->elapsed_us + last_idle_us)
- measured_us = data->elapsed_us + last_idle_us;
+ if (measured_us > data->exit_us)
+ measured_us -= data->exit_us;
+
+
+ /* update our correction ratio */
+
+ new_factor = data->correction_factor[data->bucket]
+ * (DECAY - 1) / DECAY;
+
+ if (data->expected_us > 0 && data->measured_us < MAX_INTERESTING)
+ new_factor += RESOLUTION * measured_us / data->expected_us;
else
- measured_us = -1;
+ /*
+ * we were idle so long that we count it as a perfect
+ * prediction
+ */
+ new_factor += RESOLUTION;
- /* Predict time until next break event */
- data->current_predicted_us = max(measured_us, data->last_measured_us);
+ /*
+ * We don't want 0 as factor; we always want at least
+ * a tiny bit of estimated time.
+ */
+ if (new_factor == 0)
+ new_factor = 1;
- if (last_idle_us + BREAK_FUZZ <
- data->expected_us - target->exit_latency) {
- data->last_measured_us = measured_us;
- data->elapsed_us = 0;
- } else {
- data->elapsed_us = measured_us;
- }
+ data->correction_factor[data->bucket] = new_factor;
}
/**
diff --git a/include/linux/sched.h b/include/linux/sched.h
index 17e9a8e..97b10da 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h
@@ -140,6 +140,10 @@ extern int nr_processes(void);
extern unsigned long nr_uninterruptible(void);
extern unsigned long nr_active(void);
extern unsigned long nr_iowait(void);
+extern unsigned long nr_iowait_cpu(void);
+extern unsigned long this_cpu_load(void);
+
+
extern unsigned long get_parent_ip(unsigned long addr);
diff --git a/kernel/sched.c b/kernel/sched.c
index 91843ba..0ac9053 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -2904,6 +2904,19 @@ unsigned long nr_iowait(void)
return sum;
}
+unsigned long nr_iowait_cpu(void)
+{
+ struct rq *this = this_rq();
+ return atomic_read(&this->nr_iowait);
+}
+
+unsigned long this_cpu_load(void)
+{
+ struct rq *this = this_rq();
+ return this->cpu_load[0];
+}
+
+
unsigned long nr_active(void)
{
unsigned long i, running = 0, uninterruptible = 0;
Index: kernel.spec
===================================================================
RCS file: /cvs/pkgs/rpms/kernel/F-11/kernel.spec,v
retrieving revision 1.1739
retrieving revision 1.1740
diff -u -p -r1.1739 -r1.1740
--- kernel.spec 24 Sep 2009 19:52:25 -0000 1.1739
+++ kernel.spec 24 Sep 2009 21:28:01 -0000 1.1740
@@ -702,6 +702,9 @@ Patch2904: v4l-dvb-fix-cx25840-firmware-
# fs fixes
Patch3000: linux-2.6-fs-cifs-fix-port-numbers.patch
+# ACPI
+Patch4000: linux-2.6.30-cpuidle-faster-io.patch
+
#snmp fixes
Patch10000: linux-2.6-missing-rfc2465-stats.patch
@@ -1227,6 +1230,7 @@ ApplyPatch linux-2.6-fs-cifs-fix-port-nu
# ACPI
ApplyPatch linux-2.6-defaults-acpi-video.patch
ApplyPatch linux-2.6-acpi-video-dos.patch
+ApplyPatch linux-2.6.30-cpuidle-faster-io.patch
# Various low-impact patches to aid debugging.
ApplyPatch linux-2.6-debug-sizeof-structs.patch
@@ -1971,6 +1975,10 @@ fi
# and build.
%changelog
+* Thu Sep 24 2009 Chuck Ebbert <cebbert at redhat.com> 2.6.30.8-61
+- Backport the cpuidle-faster-io fix from Fedora 12 to fix I/O
+ performance problems when reading/writing multiple disks.
+
* Thu Sep 24 2009 Chuck Ebbert <cebbert at redhat.com> 2.6.30.8-60
- Linux 2.6.30.8
More information about the fedora-extras-commits
mailing list