kernel26-alx/patches-2.6.20-r6/0003-2.6.20-sched-staircase17_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 -
 kernel/sched.c                  |   41 +++++++++++++++++++++++++++++++---------
 kernel/sysctl.c                 |    8 +++++++
 4 files changed, 52 insertions(+), 10 deletions(-)

Index: linux-2.6.20-ck1/include/linux/sched.h
===================================================================
--- linux-2.6.20-ck1.orig/include/linux/sched.h	2007-02-16 19:01:30.000000000 +1100
+++ linux-2.6.20-ck1/include/linux/sched.h	2007-02-16 19:01:30.000000000 +1100
@@ -216,7 +216,7 @@ 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-2.6.20-ck1/kernel/sched.c
===================================================================
--- linux-2.6.20-ck1.orig/kernel/sched.c	2007-02-16 19:01:30.000000000 +1100
+++ linux-2.6.20-ck1/kernel/sched.c	2007-02-16 19:01:30.000000000 +1100
@@ -63,8 +63,17 @@
 /*
  * 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 ]
@@ -96,9 +105,10 @@ 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)
 
 /*
@@ -132,6 +142,7 @@ struct rq {
 
 	/* Cached timestamp set by update_cpu_clock() */
 	unsigned long long most_recent_timestamp;
+	unsigned short cache_ticks, preempted;
 	struct task_struct *curr, *idle;
 	unsigned long next_balance;
 	struct mm_struct *prev_mm;
@@ -873,7 +884,7 @@ static inline int __normal_prio(struct 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);
@@ -1347,14 +1358,22 @@ static inline int wake_idle(int cpu, str
 #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 struct task_struct *p, struct rq *rq)
+static void fastcall preempt(const struct task_struct *p, struct rq *rq)
 {
-	if (TASK_PREEMPTS_CURR(p, rq))
-		resched_task(rq->curr);
-}
+	struct task_struct *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;
+}
 
 /***
  * try_to_wake_up - wake up a thread
@@ -3008,6 +3027,9 @@ static void task_running_tick(struct rq 
 		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);
 }
@@ -3304,6 +3326,7 @@ 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;
@@ -6625,7 +6648,7 @@ void __init sched_init(void)
 		rq = cpu_rq(i);
 		spin_lock_init(&rq->lock);
 		lockdep_set_class(&rq->lock, &rq->rq_lock_key);
-		rq->nr_running = 0;
+		rq->nr_running = rq->cache_ticks = rq->preempted = 0;
 
 #ifdef CONFIG_SMP
 		rq->sd = NULL;
Index: linux-2.6.20-ck1/kernel/sysctl.c
===================================================================
--- linux-2.6.20-ck1.orig/kernel/sysctl.c	2007-02-16 19:01:30.000000000 +1100
+++ linux-2.6.20-ck1/kernel/sysctl.c	2007-02-16 19:01:30.000000000 +1100
@@ -684,6 +684,14 @@ static ctl_table kern_table[] = {
 		.mode		= 0644,
 		.proc_handler	= &proc_dointvec,
 	},
+	{
+		.ctl_name	= CTL_UNNUMBERED,
+		.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-2.6.20-ck1/Documentation/sysctl/kernel.txt
===================================================================
--- linux-2.6.20-ck1.orig/Documentation/sysctl/kernel.txt	2007-02-16 19:01:30.000000000 +1100
+++ linux-2.6.20-ck1/Documentation/sysctl/kernel.txt	2007-02-16 19:01:30.000000000 +1100
@@ -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	1175	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			kernel/sched.c \| 41 +++++++++++++++++++++++++++++++---------
12			kernel/sysctl.c \| 8 +++++++
13			4 files changed, 52 insertions(+), 10 deletions(-)
14
15			Index: linux-2.6.20-ck1/include/linux/sched.h
16			===================================================================
17			--- linux-2.6.20-ck1.orig/include/linux/sched.h 2007-02-16 19:01:30.000000000 +1100
18			+++ linux-2.6.20-ck1/include/linux/sched.h 2007-02-16 19:01:30.000000000 +1100
19			@@ -216,7 +216,7 @@ extern void show_stack(struct task_struc
20
21			void io_schedule(void);
22			long io_schedule_timeout(long timeout);
23			-extern int sched_interactive;
24			+extern int sched_interactive, sched_compute;
25
26			extern void cpu_init (void);
27			extern void trap_init(void);
28			Index: linux-2.6.20-ck1/kernel/sched.c
29			===================================================================
30			--- linux-2.6.20-ck1.orig/kernel/sched.c 2007-02-16 19:01:30.000000000 +1100
31			+++ linux-2.6.20-ck1/kernel/sched.c 2007-02-16 19:01:30.000000000 +1100
32			@@ -63,8 +63,17 @@
33			/*
34			* sched_interactive - sysctl which allows interactive tasks to have bonus
35			* raise its priority.
36			+ * sched_compute - sysctl which enables long timeslices and delayed preemption
37			+ * for compute server usage.
38			*/
39			int sched_interactive __read_mostly = 1;
40			+int sched_compute __read_mostly;
41			+
42			+/*
43			+ * CACHE_DELAY is the time preemption is delayed in sched_compute mode
44			+ * and is set to a nominal 10ms.
45			+ */
46			+#define CACHE_DELAY (10 * (HZ) / 1001 + 1)
47
48			/*
49			* Convert user-nice values [ -20 ... 0 ... 19 ]
50			@@ -96,9 +105,10 @@ int sched_interactive __read_mostly = 1;
51
52			/*
53			* This is the time all tasks within the same priority round robin.
54			- * Set to a minimum of 6ms.
55			+ * Set to a minimum of 6ms. It is 10 times longer in compute mode.
56			*/
57			-#define RR_INTERVAL ((6 * HZ / 1001) + 1)
58			+#define _RR_INTERVAL ((6 * HZ / 1001) + 1)
59			+#define RR_INTERVAL (_RR_INTERVAL * (1 + 9 * sched_compute))
60			#define DEF_TIMESLICE (RR_INTERVAL * 19)
61
62			/*
63			@@ -132,6 +142,7 @@ struct rq {
64
65			/* Cached timestamp set by update_cpu_clock() */
66			unsigned long long most_recent_timestamp;
67			+ unsigned short cache_ticks, preempted;
68			struct task_struct curr, idle;
69			unsigned long next_balance;
70			struct mm_struct *prev_mm;
71			@@ -873,7 +884,7 @@ static inline int __normal_prio(struct t
72
73			best_bonus = bonus(p);
74			prio = MAX_RT_PRIO + best_bonus;
75			- if (sched_interactive && !batch_task(p))
76			+ if (sched_interactive && !sched_compute && !batch_task(p))
77			prio -= p->bonus;
78
79			rr = rr_interval(p);
80			@@ -1347,14 +1358,22 @@ static inline int wake_idle(int cpu, str
81			#endif
82
83			/*
84			- * Check to see if p preempts rq->curr and resched if it does.
85			+ * Check to see if p preempts rq->curr and resched if it does. In compute
86			+ * mode we do not preempt for at least CACHE_DELAY and set rq->preempted.
87			*/
88			-static inline void preempt(const struct task_struct p, struct rq rq)
89			+static void fastcall preempt(const struct task_struct p, struct rq rq)
90			{
91			- if (TASK_PREEMPTS_CURR(p, rq))
92			- resched_task(rq->curr);
93			-}
94			+ struct task_struct *curr = rq->curr;
95
96			+ if (p->prio >= curr->prio)
97			+ return;
98			+ if (!sched_compute \|\| rq->cache_ticks >= CACHE_DELAY \|\| !p->mm \|\|
99			+ rt_task(p) \|\| curr == rq->idle) {
100			+ resched_task(curr);
101			+ return;
102			+ }
103			+ rq->preempted = 1;
104			+}
105
106			/***
107			* try_to_wake_up - wake up a thread
108			@@ -3008,6 +3027,9 @@ static void task_running_tick(struct rq
109			time_slice_expired(p, rq);
110			goto out_unlock;
111			}
112			+ rq->cache_ticks++;
113			+ if (rq->preempted && rq->cache_ticks >= CACHE_DELAY)
114			+ set_tsk_need_resched(p);
115			out_unlock:
116			spin_unlock(&rq->lock);
117			}
118			@@ -3304,6 +3326,7 @@ switch_tasks:
119
120			sched_info_switch(prev, next);
121			if (likely(prev != next)) {
122			+ rq->preempted = rq->cache_ticks = 0;
123			next->timestamp = now;
124			rq->nr_switches++;
125			rq->curr = next;
126			@@ -6625,7 +6648,7 @@ void __init sched_init(void)
127			rq = cpu_rq(i);
128			spin_lock_init(&rq->lock);
129			lockdep_set_class(&rq->lock, &rq->rq_lock_key);
130			- rq->nr_running = 0;
131			+ rq->nr_running = rq->cache_ticks = rq->preempted = 0;
132
133			#ifdef CONFIG_SMP
134			rq->sd = NULL;
135			Index: linux-2.6.20-ck1/kernel/sysctl.c
136			===================================================================
137			--- linux-2.6.20-ck1.orig/kernel/sysctl.c 2007-02-16 19:01:30.000000000 +1100
138			+++ linux-2.6.20-ck1/kernel/sysctl.c 2007-02-16 19:01:30.000000000 +1100
139			@@ -684,6 +684,14 @@ static ctl_table kern_table[] = {
140			.mode = 0644,
141			.proc_handler = &proc_dointvec,
142			},
143			+ {
144			+ .ctl_name = CTL_UNNUMBERED,
145			+ .procname = "compute",
146			+ .data = &sched_compute,
147			+ .maxlen = sizeof (int),
148			+ .mode = 0644,
149			+ .proc_handler = &proc_dointvec,
150			+ },
151			#if defined(CONFIG_X86_LOCAL_APIC) && defined(CONFIG_X86)
152			{
153			.ctl_name = KERN_UNKNOWN_NMI_PANIC,
154			Index: linux-2.6.20-ck1/Documentation/sysctl/kernel.txt
155			===================================================================
156			--- linux-2.6.20-ck1.orig/Documentation/sysctl/kernel.txt 2007-02-16 19:01:30.000000000 +1100
157			+++ linux-2.6.20-ck1/Documentation/sysctl/kernel.txt 2007-02-16 19:01:30.000000000 +1100
158			@@ -18,6 +18,7 @@ Currently, these files might (depending
159			show up in /proc/sys/kernel:
160			- acpi_video_flags
161			- acct
162			+- compute
163			- core_pattern
164			- core_uses_pid
165			- ctrl-alt-del
166			@@ -85,6 +86,16 @@ valid for 30 seconds.
167
168			==============================================================
169
170			+compute:
171			+
172			+This flag controls the long timeslice, delayed preemption mode in the
173			+cpu scheduler suitable for scientific computation applications. It
174			+leads to large latencies so is unsuitable for normal usage.
175			+
176			+Disabled by default.
177			+
178			+==============================================================
179			+
180			core_pattern:
181
182			core_pattern is used to specify a core dumpfile pattern name.