Re: [PATCH]: O(1) 2.4.17-J7 Tuneable Parameters

Re: [PATCH]: O(1) 2.4.17-J7 Tuneable Parameters

[Martin Knoblauch]

> [PATCH]: O(1) 2.4.17-J7 Tuneable Parameters
> 
> 
> Hi Ingo,
> 
>         This patch synchronizes with J7 and I think makes the changes
> you wished. A couple of important points:
> 
>                 - This patch can be applied to EITHER 2.4.17 OR 2.4.18 pre 7 as
>                         long as Ingo's J7 patch is applied first.
> 
>                 - While I agree with you on not wanting these in the mainline kernel,
>                         I ran Hackbench on one of our new Foster systems with and
>                         without the tuneable parameters, and while the numbers do
>                         degrade slightly, its rather suprisingly small. So dont be afraid
>                         to use this as a system tuning aid.
> 

 How big is the actual degradation in your test? IIR, Ingo is afraid
that the tunables could easily screw things up, which of course is true.
What about adding a kernel-build option that leaves the sysctl interface
read-only by default and enables writing only if it is requested at
build time?

 That way the external interface stays constant.

Martin
-- 
------------------------------------------------------------------
Martin Knoblauch         |    email:  Martin.Knoblauch@TeraPort.de
TeraPort GmbH            |    Phone:  +49-89-510857-309
C+ITS                    |    Fax:    +49-89-510857-111
http://www.teraport.de   |    Mobile: +49-170-4904759

Re: [PATCH]: O(1) 2.4.17-J7 Tuneable Parameters

[Rob Landley]

On Wednesday 30 January 2002 08:21 am, Martin Knoblauch wrote:
> > [PATCH]: O(1) 2.4.17-J7 Tuneable Parameters
> >
> >
> > Hi Ingo,
> >
> >         This patch synchronizes with J7 and I think makes the changes
> > you wished. A couple of important points:
> >
> >                 - This patch can be applied to EITHER 2.4.17 OR 2.4.18
> > pre 7 as long as Ingo's J7 patch is applied first.
> >
> >                 - While I agree with you on not wanting these in the
> > mainline kernel, I ran Hackbench on one of our new Foster systems with
> > and without the tuneable parameters, and while the numbers do degrade
> > slightly, its rather suprisingly small. So dont be afraid to use this as
> > a system tuning aid.
>
>  How big is the actual degradation in your test? IIR, Ingo is afraid
> that the tunables could easily screw things up, which of course is true.
> What about adding a kernel-build option that leaves the sysctl interface
> read-only by default and enables writing only if it is requested at
> build time?

Nah, as long as the permissions say only root can write to it, you're pretty 
much covered.  (It's certainly no worse than root going cat /dev/zero > 
/dev/hda...)

Rob

Re: [PATCH]: O(1) 2.4.17-J7 Tuneable Parameters

[Jack F. Vogel]

On Wednesday 30 January 2002 05:21 am, Martin Knoblauch wrote:
> > [PATCH]: O(1) 2.4.17-J7 Tuneable Parameters
>
>
>  How big is the actual degradation in your test? IIR, Ingo is afraid
> that the tunables could easily screw things up, which of course is true.
> What about adding a kernel-build option that leaves the sysctl interface
> read-only by default and enables writing only if it is requested at
> build time?

Running on a machine that I dont think I can really officially give numbers..
However, lets say that without the tuneable code you got a run of hackbench
doing 60 groups that took 8.27 secs, when the tuneable code is in it went
to a whopping 8.6 secs :)

The results at least on this benchmark were all in that decimal noise.

As for a build option, if the code were integrated I might see that as 
making sense, but as this is a developmental patch the user is expected
to know what they are doing. Only root can write anything to the parameters
as well.

Cheers,

-- 
Jack F. Vogel
IBM  Linux Solutions
jfv@us.ibm.com  (work)
jfv@Bluesong.NET (home)

[PATCH]: O(1) 2.4.17-J7 Tuneable Parameters

[Jack F. Vogel]

[-- Attachment #1: Type: text/plain, Size: 618 bytes --]

Hi Ingo,

	This patch synchronizes with J7 and I think makes the changes
you wished. A couple of important points:

		- This patch can be applied to EITHER 2.4.17 OR 2.4.18 pre 7 as
			long as Ingo's J7 patch is applied first.

		- While I agree with you on not wanting these in the mainline kernel,
			I ran Hackbench on one of our new Foster systems with and
			without the tuneable parameters, and while the numbers do
			degrade slightly, its rather suprisingly small. So dont be afraid
			to use this as a system tuning aid.

-- 
Jack F. Vogel
IBM  Linux Solutions
jfv@us.ibm.com  (work)
jfv@Bluesong.NET (home)


[-- Attachment #2: Tuneable Scheduler Parameters for O(1) J7 --]
[-- Type: text/x-diff, Size: 6206 bytes --]

diff -Naur linux/include/linux/sched.h linux.jfv/include/linux/sched.h
--- linux/include/linux/sched.h	Fri Jan 25 17:31:12 2002
+++ linux.jfv/include/linux/sched.h	Fri Jan 25 15:27:13 2002
@@ -472,20 +472,37 @@
 #define DEF_USER_NICE		0
 
 /*
- * Default timeslice is 250 msecs, maximum is 500 msecs.
+ * Default timeslice is now HZ, maximum is 300 msecs.
  * Minimum timeslice is 10 msecs.
  */
-#define MIN_TIMESLICE		( 10 * HZ / 1000)
-#define MAX_TIMESLICE           (300 * HZ / 1000)
-#define CHILD_FORK_PENALTY      95
-#define PARENT_FORK_PENALTY     100
-#define EXIT_WEIGHT             3
-#define PRIO_INTERACTIVE_RATIO  20
-#define PRIO_CPU_HOG_RATIO      60
-#define PRIO_BONUS_RATIO        70
-#define INTERACTIVE_DELTA       3
-#define MAX_SLEEP_AVG           (2*HZ)
-#define STARVATION_LIMIT        (2*HZ)
+#define DEFAULT_MIN_TIMESLICE		( 10 * HZ / 1000)
+#define DEFAULT_MAX_TIMESLICE           (300 * HZ / 1000)
+#define DEFAULT_CHILD_FORK_PENALTY      95
+#define DEFAULT_PARENT_FORK_PENALTY     100
+#define DEFAULT_EXIT_WEIGHT             3
+#define DEFAULT_PRIO_INTERACTIVE_RATIO  20
+#define DEFAULT_PRIO_CPU_HOG_RATIO      60
+#define DEFAULT_PRIO_BONUS_RATIO        70
+#define DEFAULT_INTERACTIVE_DELTA       3
+#define DEFAULT_MAX_SLEEP_AVG           (2*HZ)
+#define DEFAULT_STARVATION_LIMIT        (2*HZ)
+
+extern int min_timeslice, max_timeslice, child_fork_penalty;
+extern int parent_fork_penalty, prio_cpu_hog_ratio, prio_bonus_ratio;
+extern int exit_weight, prio_bonus_ratio, interactive_delta;
+extern int max_sleep_avg, starvation_limit;
+
+#define MIN_TIMESLICE           (min_timeslice)
+#define MAX_TIMESLICE           (max_timeslice)
+#define CHILD_FORK_PENALTY      (child_fork_penalty)
+#define PARENT_FORK_PENALTY     (parent_fork_penalty)
+#define PRIO_INTERACTIVE_RATIO  (prio_interactive_ratio)
+#define PRIO_CPU_HOG_RATIO      (prio_cpu_hog_ratio)
+#define PRIO_BONUS_RATIO        (prio_bonus_ratio)
+#define INTERACTIVE_DELTA       (interactive_delta)
+#define EXIT_WEIGHT             (exit_weight)
+#define MAX_SLEEP_AVG           (max_sleep_avg)
+#define STARVATION_LIMIT        (starvation_limit)
 
 #define USER_PRIO(p)		((p)-MAX_RT_PRIO)
 #define MAX_USER_PRIO		(USER_PRIO(MAX_PRIO))
diff -Naur linux/include/linux/sysctl.h linux.jfv/include/linux/sysctl.h
--- linux/include/linux/sysctl.h	Mon Nov 26 05:29:17 2001
+++ linux.jfv/include/linux/sysctl.h	Fri Jan 25 15:28:28 2002
@@ -63,7 +63,8 @@
 	CTL_DEV=7,		/* Devices */
 	CTL_BUS=8,		/* Busses */
 	CTL_ABI=9,		/* Binary emulation */
-	CTL_CPU=10		/* CPU stuff (speed scaling, etc) */
+	CTL_CPU=10,		/* CPU stuff (speed scaling, etc) */
+	CTL_SCHED=11		/* SCHED ctl (tuneable parameters) */
 };
 
 /* CTL_BUS names: */
@@ -72,6 +73,22 @@
 	BUS_ISA=1		/* ISA */
 };
 
+/* CTL_SCHED names: */
+enum
+{
+	MAX_SLICE=1,	/* Timeslice scaling */
+	MIN_SLICE=2,
+	CHILD_PENALTY=3,
+	PARENT_PENALTY=4,
+	INT_RATIO=5,
+	HOG_RATIO=6,
+	BONUS_RATIO=7,
+	INT_DELTA=8,
+	EWEIGHT=9,
+	MAX_SLEEP=10,
+	STARVE_LIM=11
+};
+
 /* CTL_KERN names: */
 enum
 {
diff -Naur linux/kernel/sched.c linux.jfv/kernel/sched.c
--- linux/kernel/sched.c	Fri Jan 25 17:31:12 2002
+++ linux.jfv/kernel/sched.c	Fri Jan 25 15:22:40 2002
@@ -22,6 +22,22 @@
 
 #define BITMAP_SIZE ((((MAX_PRIO+7)/8)+sizeof(long)-1)/sizeof(long))
 
+/*
+**	Tuneable scheduler parameters,
+**	brought out in /proc/sys/sched
+*/
+int	max_timeslice = DEFAULT_MAX_TIMESLICE;
+int	min_timeslice = DEFAULT_MIN_TIMESLICE;
+int	child_fork_penalty = DEFAULT_CHILD_FORK_PENALTY;
+int	parent_fork_penalty = DEFAULT_PARENT_FORK_PENALTY;
+int	prio_interactive_ratio = DEFAULT_PRIO_INTERACTIVE_RATIO;
+int	prio_cpu_hog_ratio = DEFAULT_PRIO_CPU_HOG_RATIO;
+int	prio_bonus_ratio = DEFAULT_PRIO_BONUS_RATIO;
+int	interactive_delta = DEFAULT_INTERACTIVE_DELTA;
+int	exit_weight = DEFAULT_EXIT_WEIGHT;
+int	max_sleep_avg = DEFAULT_MAX_SLEEP_AVG;
+int	starvation_limit = DEFAULT_STARVATION_LIMIT;
+
 typedef struct runqueue runqueue_t;
 
 struct prio_array {
diff -Naur linux/kernel/sysctl.c linux.jfv/kernel/sysctl.c
--- linux/kernel/sysctl.c	Fri Dec 21 09:42:04 2001
+++ linux.jfv/kernel/sysctl.c	Fri Jan 25 17:10:33 2002
@@ -50,7 +50,9 @@
 extern int sysrq_enabled;
 extern int core_uses_pid;
 extern int cad_pid;
-
+extern int child_fork_penalty, parent_fork_penalty,prio_interactive_ratio;
+extern int prio_cpu_hog_ratio, prio_bonus_ratio, interactive_delta;
+extern int exit_weight, max_sleep_avg, starvation_limit;
 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and GID */
 static int maxolduid = 65535;
 static int minolduid;
@@ -109,6 +111,7 @@
 
 static ctl_table kern_table[];
 static ctl_table vm_table[];
+static ctl_table sched_table[];
 #ifdef CONFIG_NET
 extern ctl_table net_table[];
 #endif
@@ -153,6 +156,7 @@
 	{CTL_FS, "fs", NULL, 0, 0555, fs_table},
 	{CTL_DEBUG, "debug", NULL, 0, 0555, debug_table},
         {CTL_DEV, "dev", NULL, 0, 0555, dev_table},
+        {CTL_SCHED, "sched", NULL, 0, 0555, sched_table},
 	{0}
 };
 
@@ -278,6 +282,33 @@
 	{0}
 };
 
+
+static ctl_table sched_table[] = {
+	{MAX_SLICE, "MAX_TIMESLICE",
+	&max_timeslice, sizeof(int), 0644, NULL, &proc_dointvec},
+	{MIN_SLICE, "MIN_TIMESLICE",
+	&min_timeslice, sizeof(int), 0644, NULL, &proc_dointvec},
+	{CHILD_PENALTY, "CHILD_FORK_PENALTY",
+	&child_fork_penalty, sizeof(int), 0644, NULL, &proc_dointvec},
+	{PARENT_PENALTY, "PARENT_FORK_PENALTY",
+	&parent_fork_penalty, sizeof(int), 0644, NULL, &proc_dointvec},
+	{INT_RATIO, "PRIO_INTERACTIVE_RATIO",
+	&prio_interactive_ratio, sizeof(int), 0644, NULL, &proc_dointvec},
+	{HOG_RATIO, "PRIO_CPU_HOG_RATIO",
+	&prio_cpu_hog_ratio, sizeof(int), 0644, NULL, &proc_dointvec},
+	{BONUS_RATIO, "PRIO_BONUS_RATIO",
+	&prio_bonus_ratio, sizeof(int), 0644, NULL, &proc_dointvec},
+	{INT_DELTA, "INTERACTIVE_DELTA",
+	&interactive_delta, sizeof(int), 0644, NULL, &proc_dointvec},
+	{EWEIGHT, "EXIT_WEIGHT",
+	&exit_weight, sizeof(int), 0644, NULL, &proc_dointvec},
+	{MAX_SLEEP, "MAX_SLEEP_AVG",
+	&max_sleep_avg, sizeof(int), 0644, NULL, &proc_dointvec},
+	{STARVE_LIM, "INTERACTIVE_DELTA",
+	&starvation_limit, sizeof(int), 0644, NULL, &proc_dointvec},
+	{0}
+};
+
 static ctl_table proc_table[] = {
 	{0}
 };