On Wed, 2007-10-31 at 17:57 +0800, Zhang, Yanmin wrote:
quoted text
> On Tue, 2007-10-30 at 16:36 +0800, Zhang, Yanmin wrote:
> > On Tue, 2007-10-30 at 08:26 +0100, Ingo Molnar wrote:
> > > * Zhang, Yanmin <yanmin_zhang@linux.intel.com> wrote:
> > >=20
> > > > sub-bisecting captured patch=20
> > > > 38ad464d410dadceda1563f36bdb0be7fe4c8938(sched: uniform tunings)=20
> > > > caused 20% regression of aim7.
> > > >=20
> > > > The last 10% should be also related to sched parameters, such like=20
> > > > sysctl_sched_min_granularity.
> > >=20
> > > ah, interesting. Since you have CONFIG_SCHED_DEBUG enabled, could you=
=20
quoted text
> > > please try to figure out what the best value for=20
> > > /proc/sys/kernel_sched_latency, /proc/sys/kernel_sched_nr_latency and=
=20
quoted text
> > > /proc/sys/kernel_sched_min_granularity is?
> > >=20
> > > there's a tuning constraint for kernel_sched_nr_latency:=20
> > >=20
> > > - kernel_sched_nr_latency should always be set to=20
> > >   kernel_sched_latency/kernel_sched_min_granularity. (it's not a free=
=20
quoted text
> > >   tunable)
> > >=20
> > > i suspect a good approach would be to double the value of=20
> > > kernel_sched_latency and kernel_sched_nr_latency in each tuning=20
> > > iteration, while keeping kernel_sched_min_granularity unchanged. That=
=20
quoted text
> > > will excercise the tuning values of the 2.6.23 kernel as well.
> > I followed your idea to test 2.6.24-rc1. The improvement is slow.
> > When sched_nr_latency=3D2560 and sched_latency_ns=3D640000000, the perf=
ormance
quoted text
> > is still about 15% less than 2.6.23.
>=20
> I got the aim7 30% regression on my new upgraded stoakley machine. I foun=
d
quoted text
> this mahcine is slower than the old one. Maybe BIOS has issues, or memeor=
y(Might not
quoted text
> be dual-channel?) is slow. So I retested it on the old machine and found =
on the old
quoted text
> stoakley machine, the regression is about 6%, quite similiar to the regre=
ssion on tigerton
quoted text
> machine.
>=20
> By sched_nr_latency=3D640 and sched_latency_ns=3D640000000 on the old sto=
akley machine,
quoted text
> the regression becomes about 2%. Other latency has more regression.
>=20
> On my tulsa machine, by sched_nr_latency=3D640 and sched_latency_ns=3D640=
000000,
quoted text
> the regression becomes less than 1% (The original regression is about 20%=
).
quoted text
>=20
> When I ran a bad script to change the values of sched_nr_latency and sche=
d_latency_ns,
quoted text
> I hit OOPS on my tulsa machine. Below is the log. It looks like sched_nr_=
latency becomes
quoted text
> 0.

Oops, yeah I think I overlooked that case :-/
I think limiting the sysctl parameters make most sense, as a 0 value
really doesn't.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
diff --git a/kernel/sysctl.c b/kernel/sysctl.c
index 3b4efbe..0f34c91 100644
--- a/kernel/sysctl.c
+++ b/kernel/sysctl.c
@@ -94,6 +94,7 @@ static int two =3D 2;
=20
 static int zero;
 static int one_hundred =3D 100;
+static int int_max =3D INT_MAX;
=20
 /* this is needed for the proc_dointvec_minmax for [fs_]overflow UID and G=
ID */
 static int maxolduid =3D 65535;
@@ -239,7 +240,10 @@ static struct ctl_table kern_table[] =3D {
 		.data		=3D &sysctl_sched_nr_latency,
 		.maxlen		=3D sizeof(unsigned int),
 		.mode		=3D 0644,
-		.proc_handler	=3D &proc_dointvec,
+		.proc_handler	=3D &proc_dointvec_minmax,
+		.strategy	=3D &sysctl_intvec,
+		.extra1		=3D &one,
+		.extra2		=3D &int_max,
 	},
 	{
 		.ctl_name	=3D CTL_UNNUMBERED,