kernel26-alx/patches-2.6.17-r7/0005-2.6.17-sched-staircase16_compute_tunable.patch

Add the compute tunable for the staircase cpu scheduler. This modifies the
cpu scheduler behaviour for significantly longer cpu timeslices and delays
normal preemption to minimise the cpu cache harming effects of multiple
concurrent running tasks. This increases cpu throughput at the cost of
significantly increased latencies.

Signed-off-by: Con Kolivas <kernel@kolivas.org>

 Documentation/sysctl/kernel.txt |   11 +++++++++++
 include/linux/sched.h           |    2 +-
 include/linux/sysctl.h          |    1 +
 kernel/sched.c                  |   40 ++++++++++++++++++++++++++++++++--------
 kernel/sysctl.c                 |    8 ++++++++
 5 files changed, 53 insertions(+), 9 deletions(-)

Index: linux-ck-dev/include/linux/sched.h
===================================================================
--- linux-ck-dev.orig/include/linux/sched.h	2006-06-18 15:23:07.000000000 +1000
+++ linux-ck-dev/include/linux/sched.h	2006-06-18 15:23:21.000000000 +1000
@@ -202,7 +202,7 @@ extern void show_stack(struct task_struc
 
 void io_schedule(void);
 long io_schedule_timeout(long timeout);
-extern int sched_interactive;
+extern int sched_interactive, sched_compute;
 
 extern void cpu_init (void);
 extern void trap_init(void);
Index: linux-ck-dev/include/linux/sysctl.h
===================================================================
--- linux-ck-dev.orig/include/linux/sysctl.h	2006-06-18 15:23:07.000000000 +1000
+++ linux-ck-dev/include/linux/sysctl.h	2006-06-18 15:23:21.000000000 +1000
@@ -149,6 +149,7 @@ enum
 	KERN_ACPI_VIDEO_FLAGS=71, /* int: flags for setting up video after ACPI sleep */
 	KERN_IA64_UNALIGNED=72, /* int: ia64 unaligned userland trap enable */
 	KERN_INTERACTIVE=73,	/* interactive tasks can have cpu bursts */
+	KERN_COMPUTE=74,	/* adjust timeslices for a compute server */
 };
 
 
Index: linux-ck-dev/kernel/sched.c
===================================================================
--- linux-ck-dev.orig/kernel/sched.c	2006-06-18 15:23:07.000000000 +1000
+++ linux-ck-dev/kernel/sched.c	2006-06-18 15:23:21.000000000 +1000
@@ -60,8 +60,17 @@
 /*
  * sched_interactive - sysctl which allows interactive tasks to have bonus
  * raise its priority.
+ * sched_compute - sysctl which enables long timeslices and delayed preemption
+ * for compute server usage.
  */
 int sched_interactive __read_mostly = 1;
+int sched_compute __read_mostly;
+
+/*
+ * CACHE_DELAY is the time preemption is delayed in sched_compute mode
+ * and is set to a nominal 10ms.
+ */
+#define CACHE_DELAY	(10 * (HZ) / 1001 + 1)
 
 /*
  * Convert user-nice values [ -20 ... 0 ... 19 ]
@@ -92,9 +101,10 @@ int sched_interactive __read_mostly = 1;
 
 /*
  * This is the time all tasks within the same priority round robin.
- * Set to a minimum of 6ms.
+ * Set to a minimum of 6ms. It is 10 times longer in compute mode.
  */
-#define RR_INTERVAL		((6 * HZ / 1001) + 1)
+#define _RR_INTERVAL		((6 * HZ / 1001) + 1)
+#define RR_INTERVAL		(_RR_INTERVAL * (1 + 9 * sched_compute))
 #define DEF_TIMESLICE		(RR_INTERVAL * 19)
 
 #define task_hot(p, now, sd) ((long long) ((now) - (p)->timestamp)	\
@@ -135,6 +145,7 @@ struct runqueue {
 	unsigned long nr_uninterruptible;
 
 	unsigned long long timestamp_last_tick;
+	unsigned short cache_ticks, preempted;
 	task_t *curr, *idle;
 	struct mm_struct *prev_mm;
 	unsigned long bitmap[BITS_TO_LONGS(MAX_PRIO + 1)];
@@ -737,7 +748,7 @@ static int effective_prio(const task_t *
 
 	best_bonus = bonus(p);
 	prio = MAX_RT_PRIO + best_bonus;
-	if (sched_interactive && !batch_task(p))
+	if (sched_interactive && !sched_compute && !batch_task(p))
 		prio -= p->bonus;
 
 	rr = rr_interval(p);
@@ -1209,12 +1220,21 @@ static inline int wake_idle(int cpu, tas
 #endif
 
 /*
- * Check to see if p preempts rq->curr and resched if it does.
+ * Check to see if p preempts rq->curr and resched if it does. In compute
+ * mode we do not preempt for at least CACHE_DELAY and set rq->preempted.
  */
-static inline void preempt(const task_t *p, runqueue_t *rq)
+static void fastcall preempt(const task_t *p, runqueue_t *rq)
 {
-	if (TASK_PREEMPTS_CURR(p, rq))
-		resched_task(rq->curr);
+	task_t *curr = rq->curr;
+
+	if (p->prio >= curr->prio)
+		return;
+	if (!sched_compute || rq->cache_ticks >= CACHE_DELAY || !p->mm ||
+	    rt_task(p) || curr == rq->idle) {
+		resched_task(curr);
+		return;
+	}
+	rq->preempted = 1;
 }
 
 /***
@@ -2667,6 +2687,9 @@ void scheduler_tick(void)
 		time_slice_expired(p, rq);
 		goto out_unlock;
 	}
+	rq->cache_ticks++;
+	if (rq->preempted && rq->cache_ticks >= CACHE_DELAY)
+		set_tsk_need_resched(p);
 out_unlock:
 	spin_unlock(&rq->lock);
 out:
@@ -2933,6 +2956,7 @@ switch_tasks:
 
 	sched_info_switch(prev, next);
 	if (likely(prev != next)) {
+		rq->preempted = rq->cache_ticks = 0;
 		next->timestamp = now;
 		rq->nr_switches++;
 		rq->curr = next;
@@ -5971,7 +5995,7 @@ void __init sched_init(void)
 
 		rq = cpu_rq(i);
 		spin_lock_init(&rq->lock);
-		rq->nr_running = 0;
+		rq->nr_running = rq->cache_ticks = rq->preempted = 0;
 
 #ifdef CONFIG_SMP
 		rq->sd = NULL;
Index: linux-ck-dev/kernel/sysctl.c
===================================================================
--- linux-ck-dev.orig/kernel/sysctl.c	2006-06-18 15:23:07.000000000 +1000
+++ linux-ck-dev/kernel/sysctl.c	2006-06-18 15:23:21.000000000 +1000
@@ -631,6 +631,14 @@ static ctl_table kern_table[] = {
 		.mode		= 0644,
 		.proc_handler	= &proc_dointvec,
 	},
+	{
+		.ctl_name	= KERN_COMPUTE,
+		.procname	= "compute",
+		.data		= &sched_compute,
+		.maxlen		= sizeof (int),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec,
+	},
 #if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
 	{
 		.ctl_name       = KERN_UNKNOWN_NMI_PANIC,
Index: linux-ck-dev/Documentation/sysctl/kernel.txt
===================================================================
--- linux-ck-dev.orig/Documentation/sysctl/kernel.txt	2006-06-18 15:23:07.000000000 +1000
+++ linux-ck-dev/Documentation/sysctl/kernel.txt	2006-06-18 15:23:21.000000000 +1000
@@ -18,6 +18,7 @@ Currently, these files might (depending 
 show up in /proc/sys/kernel:
 - acpi_video_flags
 - acct
+- compute
 - core_pattern
 - core_uses_pid
 - ctrl-alt-del
@@ -85,6 +86,16 @@ valid for 30 seconds.
 
 ==============================================================
 
+compute:
+
+This flag controls the long timeslice, delayed preemption mode in the
+cpu scheduler suitable for scientific computation applications. It
+leads to large latencies so is unsuitable for normal usage.
+
+Disabled by default.
+
+==============================================================
+
 core_pattern:
 
 core_pattern is used to specify a core dumpfile pattern name.
1	niro	199	Add the compute tunable for the staircase cpu scheduler. This modifies the
2			cpu scheduler behaviour for significantly longer cpu timeslices and delays
3			normal preemption to minimise the cpu cache harming effects of multiple
4			concurrent running tasks. This increases cpu throughput at the cost of
5			significantly increased latencies.
6
7			Signed-off-by: Con Kolivas <kernel@kolivas.org>
8
9			Documentation/sysctl/kernel.txt \| 11 +++++++++++
10			include/linux/sched.h \| 2 +-
11			include/linux/sysctl.h \| 1 +
12			kernel/sched.c \| 40 ++++++++++++++++++++++++++++++++--------
13			kernel/sysctl.c \| 8 ++++++++
14			5 files changed, 53 insertions(+), 9 deletions(-)
15
16			Index: linux-ck-dev/include/linux/sched.h
17			===================================================================
18			--- linux-ck-dev.orig/include/linux/sched.h 2006-06-18 15:23:07.000000000 +1000
19			+++ linux-ck-dev/include/linux/sched.h 2006-06-18 15:23:21.000000000 +1000
20			@@ -202,7 +202,7 @@ extern void show_stack(struct task_struc
21
22			void io_schedule(void);
23			long io_schedule_timeout(long timeout);
24			-extern int sched_interactive;
25			+extern int sched_interactive, sched_compute;
26
27			extern void cpu_init (void);
28			extern void trap_init(void);
29			Index: linux-ck-dev/include/linux/sysctl.h
30			===================================================================
31			--- linux-ck-dev.orig/include/linux/sysctl.h 2006-06-18 15:23:07.000000000 +1000
32			+++ linux-ck-dev/include/linux/sysctl.h 2006-06-18 15:23:21.000000000 +1000
33			@@ -149,6 +149,7 @@ enum
34			KERN_ACPI_VIDEO_FLAGS=71, /* int: flags for setting up video after ACPI sleep */
35			KERN_IA64_UNALIGNED=72, /* int: ia64 unaligned userland trap enable */
36			KERN_INTERACTIVE=73, /* interactive tasks can have cpu bursts */
37			+ KERN_COMPUTE=74, /* adjust timeslices for a compute server */
38			};
39
40
41			Index: linux-ck-dev/kernel/sched.c
42			===================================================================
43			--- linux-ck-dev.orig/kernel/sched.c 2006-06-18 15:23:07.000000000 +1000
44			+++ linux-ck-dev/kernel/sched.c 2006-06-18 15:23:21.000000000 +1000
45			@@ -60,8 +60,17 @@
46			/*
47			* sched_interactive - sysctl which allows interactive tasks to have bonus
48			* raise its priority.
49			+ * sched_compute - sysctl which enables long timeslices and delayed preemption
50			+ * for compute server usage.
51			*/
52			int sched_interactive __read_mostly = 1;
53			+int sched_compute __read_mostly;
54			+
55			+/*
56			+ * CACHE_DELAY is the time preemption is delayed in sched_compute mode
57			+ * and is set to a nominal 10ms.
58			+ */
59			+#define CACHE_DELAY (10 * (HZ) / 1001 + 1)
60
61			/*
62			* Convert user-nice values [ -20 ... 0 ... 19 ]
63			@@ -92,9 +101,10 @@ int sched_interactive __read_mostly = 1;
64
65			/*
66			* This is the time all tasks within the same priority round robin.
67			- * Set to a minimum of 6ms.
68			+ * Set to a minimum of 6ms. It is 10 times longer in compute mode.
69			*/
70			-#define RR_INTERVAL ((6 * HZ / 1001) + 1)
71			+#define _RR_INTERVAL ((6 * HZ / 1001) + 1)
72			+#define RR_INTERVAL (_RR_INTERVAL * (1 + 9 * sched_compute))
73			#define DEF_TIMESLICE (RR_INTERVAL * 19)
74
75			#define task_hot(p, now, sd) ((long long) ((now) - (p)->timestamp) \
76			@@ -135,6 +145,7 @@ struct runqueue {
77			unsigned long nr_uninterruptible;
78
79			unsigned long long timestamp_last_tick;
80			+ unsigned short cache_ticks, preempted;
81			task_t curr, idle;
82			struct mm_struct *prev_mm;
83			unsigned long bitmap[BITS_TO_LONGS(MAX_PRIO + 1)];
84			@@ -737,7 +748,7 @@ static int effective_prio(const task_t *
85
86			best_bonus = bonus(p);
87			prio = MAX_RT_PRIO + best_bonus;
88			- if (sched_interactive && !batch_task(p))
89			+ if (sched_interactive && !sched_compute && !batch_task(p))
90			prio -= p->bonus;
91
92			rr = rr_interval(p);
93			@@ -1209,12 +1220,21 @@ static inline int wake_idle(int cpu, tas
94			#endif
95
96			/*
97			- * Check to see if p preempts rq->curr and resched if it does.
98			+ * Check to see if p preempts rq->curr and resched if it does. In compute
99			+ * mode we do not preempt for at least CACHE_DELAY and set rq->preempted.
100			*/
101			-static inline void preempt(const task_t p, runqueue_t rq)
102			+static void fastcall preempt(const task_t p, runqueue_t rq)
103			{
104			- if (TASK_PREEMPTS_CURR(p, rq))
105			- resched_task(rq->curr);
106			+ task_t *curr = rq->curr;
107			+
108			+ if (p->prio >= curr->prio)
109			+ return;
110			+ if (!sched_compute \|\| rq->cache_ticks >= CACHE_DELAY \|\| !p->mm \|\|
111			+ rt_task(p) \|\| curr == rq->idle) {
112			+ resched_task(curr);
113			+ return;
114			+ }
115			+ rq->preempted = 1;
116			}
117
118			/***
119			@@ -2667,6 +2687,9 @@ void scheduler_tick(void)
120			time_slice_expired(p, rq);
121			goto out_unlock;
122			}
123			+ rq->cache_ticks++;
124			+ if (rq->preempted && rq->cache_ticks >= CACHE_DELAY)
125			+ set_tsk_need_resched(p);
126			out_unlock:
127			spin_unlock(&rq->lock);
128			out:
129			@@ -2933,6 +2956,7 @@ switch_tasks:
130
131			sched_info_switch(prev, next);
132			if (likely(prev != next)) {
133			+ rq->preempted = rq->cache_ticks = 0;
134			next->timestamp = now;
135			rq->nr_switches++;
136			rq->curr = next;
137			@@ -5971,7 +5995,7 @@ void __init sched_init(void)
138
139			rq = cpu_rq(i);
140			spin_lock_init(&rq->lock);
141			- rq->nr_running = 0;
142			+ rq->nr_running = rq->cache_ticks = rq->preempted = 0;
143
144			#ifdef CONFIG_SMP
145			rq->sd = NULL;
146			Index: linux-ck-dev/kernel/sysctl.c
147			===================================================================
148			--- linux-ck-dev.orig/kernel/sysctl.c 2006-06-18 15:23:07.000000000 +1000
149			+++ linux-ck-dev/kernel/sysctl.c 2006-06-18 15:23:21.000000000 +1000
150			@@ -631,6 +631,14 @@ static ctl_table kern_table[] = {
151			.mode = 0644,
152			.proc_handler = &proc_dointvec,
153			},
154			+ {
155			+ .ctl_name = KERN_COMPUTE,
156			+ .procname = "compute",
157			+ .data = &sched_compute,
158			+ .maxlen = sizeof (int),
159			+ .mode = 0644,
160			+ .proc_handler = &proc_dointvec,
161			+ },
162			#if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
163			{
164			.ctl_name = KERN_UNKNOWN_NMI_PANIC,
165			Index: linux-ck-dev/Documentation/sysctl/kernel.txt
166			===================================================================
167			--- linux-ck-dev.orig/Documentation/sysctl/kernel.txt 2006-06-18 15:23:07.000000000 +1000
168			+++ linux-ck-dev/Documentation/sysctl/kernel.txt 2006-06-18 15:23:21.000000000 +1000
169			@@ -18,6 +18,7 @@ Currently, these files might (depending
170			show up in /proc/sys/kernel:
171			- acpi_video_flags
172			- acct
173			+- compute
174			- core_pattern
175			- core_uses_pid
176			- ctrl-alt-del
177			@@ -85,6 +86,16 @@ valid for 30 seconds.
178
179			==============================================================
180
181			+compute:
182			+
183			+This flag controls the long timeslice, delayed preemption mode in the
184			+cpu scheduler suitable for scientific computation applications. It
185			+leads to large latencies so is unsuitable for normal usage.
186			+
187			+Disabled by default.
188			+
189			+==============================================================
190			+
191			core_pattern:
192
193			core_pattern is used to specify a core dumpfile pattern name.