Hi, Folks
Charles has raised a problem that we don't have any tool yet
for testing the scheduler with out any disturb from other
subsystem, and I also found it's hard to test scheduler optimize
patch, since the improvement could be easily eaten by other
subsystem like IO.
So Let's check the tools we have currently:
1. perf sched
we can use it to trace the threads we interested, and
the info it provided is very good, but one issue is,
it could not create the workload we want, also collect
the info and do summary is not so easy.
2. linsched
It's a very good tool to create the test environment,
but it's implementation is to ideal, so it could not
present the real world problem.
Since both perf and linsched could not meet our requirement, we
decided to develop a new tool, let's currently call it
scheduler test module(STM).
It's propose is:
1. create the workload we want.
2. test the pure scheduler.
3. collect info we need and do summary.
This tool should be very easy to use and not depends on
the implementation of scheduler.
We can use it to check the pure scheduler performance on
our system.
We can use it to check whether there are regression in
scheduler when testing patches.
And other usage I've not figure out yet.
In order to explain the idea more directly, I have wrote a
prototype STM, it's a separate module, and you can use it
just like 'rcutorture'.
I attached a small script 'play.sh' to help you easily
run the test, put 'schedtm.c' and 'play.sh' in same directory
and run 'play.sh', you will see out put like:
schedtm: summary
schedtm: cpu count: cpu run preempt
schedtm: 0 1381 1381
schedtm: 1 957 955
schedtm: 2 900 900
schedtm: 3 1035 1034
schedtm: 4 991 990
schedtm: 5 940 939
schedtm: 6 900 897
schedtm: 7 942 948
schedtm: 8 852 850
schedtm: 9 931 938
schedtm: 10 936 934
schedtm: 11 951 950
schedtm: total time(us): 10138172
schedtm: run time(us): 5055223(49.86%)
schedtm: wait time(us): 5082949
schedtm: latency(us): 10489
schedtm: stmt22 got highest run time 5604941(+10%)
schedtm: stmt3 got lowest run time 4852057(-4%)
schedtm: stmt12 got highest latency 11482(+9%)
schedtm: stmt0 got lowest latency 7561(-27%)
And you can enable/disable CONFIG_PREEMPT to see magnificent
change on latency.
This is nothing but a demo, and please "RUN IT ON A TEST MACHINE"...
It will create 24 kernel threads and run 10 seconds, you can change
it by module param.
I will be appreciate if I could get some feedback from the scheduler
experts like you, whatever you think it's good or junk, please let
me know :)
Regards,
Michael Wang
play.sh:
DURATION=10
NORMAL_THREADS=24
PERIOD=10
make clean
make
insmod ./schedtm.ko normalnr=$NORMAL_THREADS period=$PERIOD
sleep $DURATION
rmmod ./schedtm.ko
dmesg | grep schedtm
schedtm.c:
/*
* scheduler test module
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation; either version 2 of the License, or
* (at your option) any later version.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
*
* Copyright (C) IBM Corporation, 2012
*
* Authors: Michael Wang <[email protected]>
*
*/
#include <linux/module.h>
#include <linux/kthread.h>
#include <linux/slab.h>
MODULE_LICENSE("GPL");
MODULE_AUTHOR("Michael Wang <[email protected]>");
#define pr_schedtm(fmt, ...) \
do { \
pr_alert("schedtm: "); \
printk(pr_fmt(fmt), ##__VA_ARGS__); \
} while (0)
struct schedtm_data {
struct task_struct *t;
unsigned long run_count[NR_CPUS];
unsigned long preempt_count[NR_CPUS];
unsigned long start_time;
unsigned long total_time;
unsigned long wait_time;
unsigned long run_time;
unsigned long run_time_tmp;
unsigned long latency;
unsigned long latency_tmp;
unsigned long latency_count;
struct list_head list;
struct preempt_notifier pn;
};
LIST_HEAD(schedtm_data_list);
static struct schedtm_data *highest_run_time;
static struct schedtm_data *lowest_run_time;
static struct schedtm_data *highest_latency;
static struct schedtm_data *lowest_latency;
static unsigned int normalnr = 1;
static unsigned int period = 10;
module_param(normalnr, uint, 0444);
MODULE_PARM_DESC(normalnr, "Number of normal threads");
module_param(period, uint, 0444);
MODULE_PARM_DESC(period, "Running period in jiffies");
void schedtm_sched_in(struct preempt_notifier *pn, int cpu)
{
struct schedtm_data *data = container_of(pn, struct schedtm_data, pn);
WARN_ON(data->t != current);
data->run_time_tmp = ktime_to_us(ktime_get());
data->latency += ktime_to_us(ktime_get()) - data->latency_tmp;
data->latency_count++;
data->run_count[cpu]++;
}
void schedtm_sched_out(struct preempt_notifier *pn, struct task_struct *next)
{
int cpu = smp_processor_id();
struct schedtm_data *data = container_of(pn, struct schedtm_data, pn);
WARN_ON(data->t != current);
data->run_time += ktime_to_us(ktime_get()) - data->run_time_tmp;
data->latency_tmp = ktime_to_us(ktime_get());
data->preempt_count[cpu]++;
}
struct preempt_ops schedtm_po = {
.sched_in = schedtm_sched_in,
.sched_out = schedtm_sched_out,
};
static int schedtm_thread_fn(void *arg)
{
struct schedtm_data *data = (struct schedtm_data *)arg;
unsigned long checkp = jiffies;
data->start_time = ktime_to_us(ktime_get());
data->run_time_tmp = ktime_to_us(ktime_get());
data->pn.ops = &schedtm_po;
preempt_notifier_register(&data->pn);
do {
WARN_ON(current != data->t);
if (jiffies - checkp > period) {
checkp = jiffies;
schedule();
}
} while (!kthread_should_stop());
preempt_notifier_unregister(&data->pn);
data->total_time = ktime_to_us(ktime_get()) - data->start_time;
data->wait_time = data->total_time - data->run_time;
data->latency = data->latency / (data->latency_count + 1);
return 0;
}
static void schedtm_report_cpu(struct schedtm_data *data)
{
int cpu = 0;
pr_schedtm("\tcpu count:\tcpu\trun\tpreempt\n");
for_each_online_cpu(cpu) {
pr_schedtm("\t\t\t%d\t%ld\t%ld\n", cpu,
data->run_count[cpu],
data->preempt_count[cpu]);
}
}
static void schedtm_print_time(unsigned long us)
{
/*
unsigned long s = us / 1000000;
unsigned long ms = (us - s * 1000000) / 1000;
us = us - s * 1000000 - ms * 1000;
pr_schedtm("\t\t\t%lds%ldms%ldus\n", s, ms, us);
*/
}
static void schedtm_report_run(struct schedtm_data *data)
{
unsigned long rp = (10000 * data->run_time) / (data->total_time + 1);
unsigned long rpa = rp / 100;
unsigned long rpb = rp - rpa * 100;
pr_schedtm("\ttotal time(us):\t%ld\n", data->total_time);
schedtm_print_time(data->total_time);
pr_schedtm("\trun time(us):\t%ld(%ld.%ld%%)\n",
data->run_time, rpa, rpb);
schedtm_print_time(data->run_time);
pr_schedtm("\twait time(us):\t%ld\n", data->wait_time);
schedtm_print_time(data->wait_time);
}
static void schedtm_report_latency(struct schedtm_data *data)
{
pr_schedtm("\tlatency(us):\t%ld\n", data->latency);
schedtm_print_time(data->latency);
}
static void schedtm_report_all(void)
{
/*
struct schedtm_data *data = NULL;
list_for_each_entry(data, &schedtm_data_list, list) {
pr_schedtm("%s\n", data->t->comm);
schedtm_report_cpu(data);
schedtm_report_run(data);
schedtm_report_latency(data);
}
*/
}
static void schedtm_summary_calculate(struct schedtm_data *summary)
{
int cpu = 0;
struct schedtm_data *data = NULL;
list_for_each_entry(data, &schedtm_data_list, list) {
for_each_online_cpu(cpu) {
summary->run_count[cpu] += data->run_count[cpu];
summary->preempt_count[cpu] += data->preempt_count[cpu];
}
summary->total_time += data->total_time;
summary->run_time += data->run_time;
summary->wait_time += data->wait_time;
summary->latency += data->latency;
if (!highest_run_time ||
highest_run_time->run_time < data->run_time)
highest_run_time = data;
if (!lowest_run_time ||
lowest_run_time->run_time > data->run_time)
lowest_run_time = data;
if (!highest_latency ||
highest_latency->latency < data->latency)
highest_latency = data;
if (!lowest_latency ||
lowest_latency->latency > data->latency)
lowest_latency = data;
}
summary->total_time /= normalnr;
summary->run_time /= normalnr;
summary->wait_time /= normalnr;
summary->latency /= normalnr;
}
static void schedtm_report_special(struct schedtm_data *summary)
{
pr_schedtm("%s got highest run time %ld(+%ld%%)\n",
highest_run_time->t->comm,
highest_run_time->run_time,
(100 * (highest_run_time->run_time - summary->run_time))
/ (summary->run_time + 1));
pr_schedtm("%s got lowest run time %ld(-%ld%%)\n",
lowest_run_time->t->comm,
lowest_run_time->run_time,
(100 * (summary->run_time - lowest_run_time->run_time))
/ (summary->run_time + 1));
pr_schedtm("%s got highest latency %ld(+%ld%%)\n",
highest_latency->t->comm,
highest_latency->latency,
(100 * (highest_latency->latency - summary->latency))
/ (summary->latency + 1));
pr_schedtm("%s got lowest latency %ld(-%ld%%)\n",
lowest_latency->t->comm,
lowest_latency->latency,
(100 * (summary->latency - lowest_latency->latency))
/ (summary->latency + 1));
}
static void schedtm_summary_all(void)
{
struct schedtm_data *summary =
kzalloc(sizeof(struct schedtm_data), GFP_KERNEL);
if (!summary) {
pr_schedtm("failed to summary\n");
return;
}
schedtm_summary_calculate(summary);
pr_schedtm("summary\n");
schedtm_report_cpu(summary);
schedtm_report_run(summary);
schedtm_report_latency(summary);
schedtm_report_special(summary);
kfree(summary);
}
static int schedtm_succeed(void)
{
int threadnr = 0;
struct schedtm_data *data = NULL;
list_for_each_entry(data, &schedtm_data_list, list) {
kthread_stop(data->t);
threadnr++;
}
return threadnr == normalnr;
}
static void schedtm_exit(void)
{
struct schedtm_data *data = NULL;
if (schedtm_succeed()) {
schedtm_report_all();
schedtm_summary_all();
}
list_for_each_entry(data, &schedtm_data_list, list)
kfree(data);
pr_schedtm("scheduler test module off\n");
}
static int __init schedtm_check_param(void)
{
if (normalnr == 0) {
pr_schedtm("illegal normalnr %d\n", normalnr);
return 0;
}
if (normalnr >= nr_cpu_ids * 100) {
pr_schedtm("too big normalnr %d for %d cpu system\n",
normalnr, nr_cpu_ids);
return 0;
}
#ifndef CONFIG_PREEMPT
if (period >= 100) {
pr_schedtm("too big period %d for none kernel-preempt system\n",
period);
return 0;
}
#endif
return 1;
}
static int __init schedtm_init(void)
{
int threadnr = normalnr;
struct schedtm_data *data = NULL;
if (!schedtm_check_param())
return -EAGAIN;
pr_schedtm("scheduler test module on\n");
while (threadnr--) {
data = kzalloc(sizeof(struct schedtm_data), GFP_KERNEL);
if (!data)
goto failed;
data->t = kthread_create(schedtm_thread_fn, data,
"stmt%d", threadnr);
if (!data->t) {
kfree(data);
goto failed;
}
list_add_tail(&data->list, &schedtm_data_list);
}
list_for_each_entry(data, &schedtm_data_list, list)
wake_up_process(data->t);
return 0;
failed:
list_for_each_entry(data, &schedtm_data_list, list) {
kthread_stop(data->t);
kfree(data);
}
return -ENOMEM;
}
module_init(schedtm_init);
module_exit(schedtm_exit);
Yes, it's a new way to do scheduler test. But why use kernel
threads? The info, total time, run time, wait time, preempt number,
all can be collected from tasks' sched info from /proc/pid/sched and
/proc/pid/stat.
I don't understand clearly about "pure scheduler performance" here. In
order to test scheduler fully, we need to do IO test, and other
subsystems will be involved in. But if other subsystems' code are not
changed, the result still can refer to scheduler's change.
I can't think much farther about the advantage this way could give.
Maybe you should show us ur better examples. :)
Regards,
Charles
On 10/25/2012 01:40 PM, Michael Wang wrote:
> Hi, Folks
>
> Charles has raised a problem that we don't have any tool yet
> for testing the scheduler with out any disturb from other
> subsystem, and I also found it's hard to test scheduler optimize
> patch, since the improvement could be easily eaten by other
> subsystem like IO.
>
> So Let's check the tools we have currently:
> 1. perf sched
>
> we can use it to trace the threads we interested, and
> the info it provided is very good, but one issue is,
> it could not create the workload we want, also collect
> the info and do summary is not so easy.
>
> 2. linsched
>
> It's a very good tool to create the test environment,
> but it's implementation is to ideal, so it could not
> present the real world problem.
>
> Since both perf and linsched could not meet our requirement, we
> decided to develop a new tool, let's currently call it
> scheduler test module(STM).
>
> It's propose is:
> 1. create the workload we want.
> 2. test the pure scheduler.
> 3. collect info we need and do summary.
>
> This tool should be very easy to use and not depends on
> the implementation of scheduler.
>
> We can use it to check the pure scheduler performance on
> our system.
>
> We can use it to check whether there are regression in
> scheduler when testing patches.
>
> And other usage I've not figure out yet.
>
> In order to explain the idea more directly, I have wrote a
> prototype STM, it's a separate module, and you can use it
> just like 'rcutorture'.
>
> I attached a small script 'play.sh' to help you easily
> run the test, put 'schedtm.c' and 'play.sh' in same directory
> and run 'play.sh', you will see out put like:
>
> schedtm: summary
> schedtm: cpu count: cpu run preempt
> schedtm: 0 1381 1381
> schedtm: 1 957 955
> schedtm: 2 900 900
> schedtm: 3 1035 1034
> schedtm: 4 991 990
> schedtm: 5 940 939
> schedtm: 6 900 897
> schedtm: 7 942 948
> schedtm: 8 852 850
> schedtm: 9 931 938
> schedtm: 10 936 934
> schedtm: 11 951 950
> schedtm: total time(us): 10138172
> schedtm: run time(us): 5055223(49.86%)
> schedtm: wait time(us): 5082949
> schedtm: latency(us): 10489
> schedtm: stmt22 got highest run time 5604941(+10%)
> schedtm: stmt3 got lowest run time 4852057(-4%)
> schedtm: stmt12 got highest latency 11482(+9%)
> schedtm: stmt0 got lowest latency 7561(-27%)
>
> And you can enable/disable CONFIG_PREEMPT to see magnificent
> change on latency.
>
> This is nothing but a demo, and please "RUN IT ON A TEST MACHINE"...
>
> It will create 24 kernel threads and run 10 seconds, you can change
> it by module param.
>
> I will be appreciate if I could get some feedback from the scheduler
> experts like you, whatever you think it's good or junk, please let
> me know :)
>
> Regards,
> Michael Wang
>
>
>
> play.sh:
>
> DURATION=10
> NORMAL_THREADS=24
> PERIOD=10
>
> make clean
> make
> insmod ./schedtm.ko normalnr=$NORMAL_THREADS period=$PERIOD
> sleep $DURATION
> rmmod ./schedtm.ko
> dmesg | grep schedtm
>
> schedtm.c:
>
> /*
> * scheduler test module
> *
> * This program is free software; you can redistribute it and/or modify
> * it under the terms of the GNU General Public License as published by
> * the Free Software Foundation; either version 2 of the License, or
> * (at your option) any later version.
> *
> * This program is distributed in the hope that it will be useful,
> * but WITHOUT ANY WARRANTY; without even the implied warranty of
> * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
> * GNU General Public License for more details.
> *
> * You should have received a copy of the GNU General Public License
> * along with this program; if not, write to the Free Software
> * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
> *
> * Copyright (C) IBM Corporation, 2012
> *
> * Authors: Michael Wang <[email protected]>
> *
> */
> #include <linux/module.h>
> #include <linux/kthread.h>
> #include <linux/slab.h>
>
> MODULE_LICENSE("GPL");
> MODULE_AUTHOR("Michael Wang <[email protected]>");
>
> #define pr_schedtm(fmt, ...) \
> do { \
> pr_alert("schedtm: "); \
> printk(pr_fmt(fmt), ##__VA_ARGS__); \
> } while (0)
>
> struct schedtm_data {
> struct task_struct *t;
> unsigned long run_count[NR_CPUS];
> unsigned long preempt_count[NR_CPUS];
> unsigned long start_time;
> unsigned long total_time;
> unsigned long wait_time;
> unsigned long run_time;
> unsigned long run_time_tmp;
> unsigned long latency;
> unsigned long latency_tmp;
> unsigned long latency_count;
> struct list_head list;
> struct preempt_notifier pn;
> };
>
> LIST_HEAD(schedtm_data_list);
> static struct schedtm_data *highest_run_time;
> static struct schedtm_data *lowest_run_time;
> static struct schedtm_data *highest_latency;
> static struct schedtm_data *lowest_latency;
>
> static unsigned int normalnr = 1;
> static unsigned int period = 10;
>
> module_param(normalnr, uint, 0444);
> MODULE_PARM_DESC(normalnr, "Number of normal threads");
>
> module_param(period, uint, 0444);
> MODULE_PARM_DESC(period, "Running period in jiffies");
>
> void schedtm_sched_in(struct preempt_notifier *pn, int cpu)
> {
> struct schedtm_data *data = container_of(pn, struct schedtm_data, pn);
> WARN_ON(data->t != current);
>
> data->run_time_tmp = ktime_to_us(ktime_get());
> data->latency += ktime_to_us(ktime_get()) - data->latency_tmp;
> data->latency_count++;
> data->run_count[cpu]++;
> }
>
> void schedtm_sched_out(struct preempt_notifier *pn, struct task_struct *next)
> {
> int cpu = smp_processor_id();
> struct schedtm_data *data = container_of(pn, struct schedtm_data, pn);
> WARN_ON(data->t != current);
>
> data->run_time += ktime_to_us(ktime_get()) - data->run_time_tmp;
> data->latency_tmp = ktime_to_us(ktime_get());
> data->preempt_count[cpu]++;
> }
>
> struct preempt_ops schedtm_po = {
> .sched_in = schedtm_sched_in,
> .sched_out = schedtm_sched_out,
> };
>
> static int schedtm_thread_fn(void *arg)
> {
> struct schedtm_data *data = (struct schedtm_data *)arg;
> unsigned long checkp = jiffies;
>
> data->start_time = ktime_to_us(ktime_get());
> data->run_time_tmp = ktime_to_us(ktime_get());
> data->pn.ops = &schedtm_po;
> preempt_notifier_register(&data->pn);
>
> do {
> WARN_ON(current != data->t);
> if (jiffies - checkp > period) {
> checkp = jiffies;
> schedule();
> }
> } while (!kthread_should_stop());
>
> preempt_notifier_unregister(&data->pn);
> data->total_time = ktime_to_us(ktime_get()) - data->start_time;
> data->wait_time = data->total_time - data->run_time;
> data->latency = data->latency / (data->latency_count + 1);
>
> return 0;
> }
>
> static void schedtm_report_cpu(struct schedtm_data *data)
> {
> int cpu = 0;
>
> pr_schedtm("\tcpu count:\tcpu\trun\tpreempt\n");
>
> for_each_online_cpu(cpu) {
> pr_schedtm("\t\t\t%d\t%ld\t%ld\n", cpu,
> data->run_count[cpu],
> data->preempt_count[cpu]);
> }
> }
>
> static void schedtm_print_time(unsigned long us)
> {
> /*
> unsigned long s = us / 1000000;
> unsigned long ms = (us - s * 1000000) / 1000;
> us = us - s * 1000000 - ms * 1000;
> pr_schedtm("\t\t\t%lds%ldms%ldus\n", s, ms, us);
> */
> }
>
> static void schedtm_report_run(struct schedtm_data *data)
> {
> unsigned long rp = (10000 * data->run_time) / (data->total_time + 1);
> unsigned long rpa = rp / 100;
> unsigned long rpb = rp - rpa * 100;
>
> pr_schedtm("\ttotal time(us):\t%ld\n", data->total_time);
> schedtm_print_time(data->total_time);
>
> pr_schedtm("\trun time(us):\t%ld(%ld.%ld%%)\n",
> data->run_time, rpa, rpb);
> schedtm_print_time(data->run_time);
>
> pr_schedtm("\twait time(us):\t%ld\n", data->wait_time);
> schedtm_print_time(data->wait_time);
> }
>
> static void schedtm_report_latency(struct schedtm_data *data)
> {
> pr_schedtm("\tlatency(us):\t%ld\n", data->latency);
> schedtm_print_time(data->latency);
> }
>
> static void schedtm_report_all(void)
> {
> /*
> struct schedtm_data *data = NULL;
> list_for_each_entry(data, &schedtm_data_list, list) {
> pr_schedtm("%s\n", data->t->comm);
>
> schedtm_report_cpu(data);
> schedtm_report_run(data);
> schedtm_report_latency(data);
> }
> */
> }
>
> static void schedtm_summary_calculate(struct schedtm_data *summary)
> {
> int cpu = 0;
> struct schedtm_data *data = NULL;
>
> list_for_each_entry(data, &schedtm_data_list, list) {
> for_each_online_cpu(cpu) {
> summary->run_count[cpu] += data->run_count[cpu];
> summary->preempt_count[cpu] += data->preempt_count[cpu];
> }
>
> summary->total_time += data->total_time;
> summary->run_time += data->run_time;
> summary->wait_time += data->wait_time;
> summary->latency += data->latency;
>
> if (!highest_run_time ||
> highest_run_time->run_time < data->run_time)
> highest_run_time = data;
> if (!lowest_run_time ||
> lowest_run_time->run_time > data->run_time)
> lowest_run_time = data;
> if (!highest_latency ||
> highest_latency->latency < data->latency)
> highest_latency = data;
> if (!lowest_latency ||
> lowest_latency->latency > data->latency)
> lowest_latency = data;
> }
>
> summary->total_time /= normalnr;
> summary->run_time /= normalnr;
> summary->wait_time /= normalnr;
> summary->latency /= normalnr;
> }
>
> static void schedtm_report_special(struct schedtm_data *summary)
> {
> pr_schedtm("%s got highest run time %ld(+%ld%%)\n",
> highest_run_time->t->comm,
> highest_run_time->run_time,
> (100 * (highest_run_time->run_time - summary->run_time))
> / (summary->run_time + 1));
>
> pr_schedtm("%s got lowest run time %ld(-%ld%%)\n",
> lowest_run_time->t->comm,
> lowest_run_time->run_time,
> (100 * (summary->run_time - lowest_run_time->run_time))
> / (summary->run_time + 1));
>
> pr_schedtm("%s got highest latency %ld(+%ld%%)\n",
> highest_latency->t->comm,
> highest_latency->latency,
> (100 * (highest_latency->latency - summary->latency))
> / (summary->latency + 1));
>
> pr_schedtm("%s got lowest latency %ld(-%ld%%)\n",
> lowest_latency->t->comm,
> lowest_latency->latency,
> (100 * (summary->latency - lowest_latency->latency))
> / (summary->latency + 1));
> }
>
> static void schedtm_summary_all(void)
> {
> struct schedtm_data *summary =
> kzalloc(sizeof(struct schedtm_data), GFP_KERNEL);
> if (!summary) {
> pr_schedtm("failed to summary\n");
> return;
> }
>
> schedtm_summary_calculate(summary);
>
> pr_schedtm("summary\n");
>
> schedtm_report_cpu(summary);
> schedtm_report_run(summary);
> schedtm_report_latency(summary);
> schedtm_report_special(summary);
>
> kfree(summary);
> }
>
> static int schedtm_succeed(void)
> {
> int threadnr = 0;
> struct schedtm_data *data = NULL;
>
> list_for_each_entry(data, &schedtm_data_list, list) {
> kthread_stop(data->t);
> threadnr++;
> }
>
> return threadnr == normalnr;
> }
>
> static void schedtm_exit(void)
> {
> struct schedtm_data *data = NULL;
>
> if (schedtm_succeed()) {
> schedtm_report_all();
> schedtm_summary_all();
> }
>
> list_for_each_entry(data, &schedtm_data_list, list)
> kfree(data);
>
> pr_schedtm("scheduler test module off\n");
> }
>
> static int __init schedtm_check_param(void)
> {
> if (normalnr == 0) {
> pr_schedtm("illegal normalnr %d\n", normalnr);
> return 0;
> }
>
> if (normalnr >= nr_cpu_ids * 100) {
> pr_schedtm("too big normalnr %d for %d cpu system\n",
> normalnr, nr_cpu_ids);
> return 0;
> }
>
> #ifndef CONFIG_PREEMPT
> if (period >= 100) {
> pr_schedtm("too big period %d for none kernel-preempt system\n",
> period);
> return 0;
> }
> #endif
>
> return 1;
> }
>
> static int __init schedtm_init(void)
> {
> int threadnr = normalnr;
> struct schedtm_data *data = NULL;
>
> if (!schedtm_check_param())
> return -EAGAIN;
>
> pr_schedtm("scheduler test module on\n");
>
> while (threadnr--) {
> data = kzalloc(sizeof(struct schedtm_data), GFP_KERNEL);
> if (!data)
> goto failed;
>
> data->t = kthread_create(schedtm_thread_fn, data,
> "stmt%d", threadnr);
> if (!data->t) {
> kfree(data);
> goto failed;
> }
>
> list_add_tail(&data->list, &schedtm_data_list);
> }
>
> list_for_each_entry(data, &schedtm_data_list, list)
> wake_up_process(data->t);
>
> return 0;
>
> failed:
> list_for_each_entry(data, &schedtm_data_list, list) {
> kthread_stop(data->t);
> kfree(data);
> }
>
> return -ENOMEM;
> }
>
> module_init(schedtm_init);
> module_exit(schedtm_exit);
>
On 10/26/2012 10:27 AM, Charles Wang wrote:
> Yes, it's a new way to do scheduler test. But why use kernel
> threads? The info, total time, run time, wait time, preempt number, all
> can be collected from tasks' sched info from /proc/pid/sched and
> /proc/pid/stat.
That's right, next version's STM could get info like proc, actually
there are many ways to implement it, but we need to figure out whether
STM is needed firstly ;-)
>
> I don't understand clearly about "pure scheduler performance" here. In
> order to test scheduler fully, we need to do IO test, and other
> subsystems will be involved in. But if other subsystems' code are not
> changed, the result still can refer to scheduler's change.
For example when we test the latency of net traffic, there are software,
nic driver, nic, router, and the same thing on peer machine, all those
staff besides scheduler are unstable even the code not changed, and can
easily influence the test result.
Let's say we changed some param like 'sched_latency_ns' which we think
may help to reduce the latency of system, and suppose it really helped
to reduce 0.xx% latency, how can we detect that while so many unstable
thing in our test?
Then after we have done many changes, suppose each of them can help
increase 0.xx% but we can't know, so we just think those param are
useless, but actually the accumulate improvement may be x% which we
really care.
So the pure means with out any influence from other thing.
But I think STM should also have the ability to do the normal test which
included all the subsystem we want, and then we could know which one is
the real bottleneck.
>
> I can't think much farther about the advantage this way could give.
> Maybe you should show us ur better examples. :)
I'd like to make it more useful not just a demo, but I need more
feedback and suggestions :)
Regards,
Michael Wang
>
> Regards,
> Charles
>
> On 10/25/2012 01:40 PM, Michael Wang wrote:
>> Hi, Folks
>>
>> Charles has raised a problem that we don't have any tool yet
>> for testing the scheduler with out any disturb from other
>> subsystem, and I also found it's hard to test scheduler optimize
>> patch, since the improvement could be easily eaten by other
>> subsystem like IO.
>>
>> So Let's check the tools we have currently:
>> 1. perf sched
>>
>> we can use it to trace the threads we interested, and
>> the info it provided is very good, but one issue is,
>> it could not create the workload we want, also collect
>> the info and do summary is not so easy.
>>
>> 2. linsched
>>
>> It's a very good tool to create the test environment,
>> but it's implementation is to ideal, so it could not
>> present the real world problem.
>>
>> Since both perf and linsched could not meet our requirement, we
>> decided to develop a new tool, let's currently call it
>> scheduler test module(STM).
>>
>> It's propose is:
>> 1. create the workload we want.
>> 2. test the pure scheduler.
>> 3. collect info we need and do summary.
>>
>> This tool should be very easy to use and not depends on
>> the implementation of scheduler.
>>
>> We can use it to check the pure scheduler performance on
>> our system.
>>
>> We can use it to check whether there are regression in
>> scheduler when testing patches.
>>
>> And other usage I've not figure out yet.
>>
>> In order to explain the idea more directly, I have wrote a
>> prototype STM, it's a separate module, and you can use it
>> just like 'rcutorture'.
>>
>> I attached a small script 'play.sh' to help you easily
>> run the test, put 'schedtm.c' and 'play.sh' in same directory
>> and run 'play.sh', you will see out put like:
>>
>> schedtm: summary
>> schedtm: cpu count: cpu run preempt
>> schedtm: 0 1381 1381
>> schedtm: 1 957 955
>> schedtm: 2 900 900
>> schedtm: 3 1035 1034
>> schedtm: 4 991 990
>> schedtm: 5 940 939
>> schedtm: 6 900 897
>> schedtm: 7 942 948
>> schedtm: 8 852 850
>> schedtm: 9 931 938
>> schedtm: 10 936 934
>> schedtm: 11 951 950
>> schedtm: total time(us): 10138172
>> schedtm: run time(us): 5055223(49.86%)
>> schedtm: wait time(us): 5082949
>> schedtm: latency(us): 10489
>> schedtm: stmt22 got highest run time 5604941(+10%)
>> schedtm: stmt3 got lowest run time 4852057(-4%)
>> schedtm: stmt12 got highest latency 11482(+9%)
>> schedtm: stmt0 got lowest latency 7561(-27%)
>>
>> And you can enable/disable CONFIG_PREEMPT to see magnificent
>> change on latency.
>>
>> This is nothing but a demo, and please "RUN IT ON A TEST MACHINE"...
>>
>> It will create 24 kernel threads and run 10 seconds, you can change
>> it by module param.
>>
>> I will be appreciate if I could get some feedback from the scheduler
>> experts like you, whatever you think it's good or junk, please let
>> me know :)
>>
>> Regards,
>> Michael Wang
>>
>>
>>
>> play.sh:
>>
>> DURATION=10
>> NORMAL_THREADS=24
>> PERIOD=10
>>
>> make clean
>> make
>> insmod ./schedtm.ko normalnr=$NORMAL_THREADS period=$PERIOD
>> sleep $DURATION
>> rmmod ./schedtm.ko
>> dmesg | grep schedtm
>>
>> schedtm.c:
>>
>> /*
>> * scheduler test module
>> *
>> * This program is free software; you can redistribute it and/or modify
>> * it under the terms of the GNU General Public License as published by
>> * the Free Software Foundation; either version 2 of the License, or
>> * (at your option) any later version.
>> *
>> * This program is distributed in the hope that it will be useful,
>> * but WITHOUT ANY WARRANTY; without even the implied warranty of
>> * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
>> * GNU General Public License for more details.
>> *
>> * You should have received a copy of the GNU General Public License
>> * along with this program; if not, write to the Free Software
>> * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA
>> 02111-1307, USA.
>> *
>> * Copyright (C) IBM Corporation, 2012
>> *
>> * Authors: Michael Wang <[email protected]>
>> *
>> */
>> #include <linux/module.h>
>> #include <linux/kthread.h>
>> #include <linux/slab.h>
>>
>> MODULE_LICENSE("GPL");
>> MODULE_AUTHOR("Michael Wang <[email protected]>");
>>
>> #define pr_schedtm(fmt, ...) \
>> do { \
>> pr_alert("schedtm: "); \
>> printk(pr_fmt(fmt), ##__VA_ARGS__); \
>> } while (0)
>>
>> struct schedtm_data {
>> struct task_struct *t;
>> unsigned long run_count[NR_CPUS];
>> unsigned long preempt_count[NR_CPUS];
>> unsigned long start_time;
>> unsigned long total_time;
>> unsigned long wait_time;
>> unsigned long run_time;
>> unsigned long run_time_tmp;
>> unsigned long latency;
>> unsigned long latency_tmp;
>> unsigned long latency_count;
>> struct list_head list;
>> struct preempt_notifier pn;
>> };
>>
>> LIST_HEAD(schedtm_data_list);
>> static struct schedtm_data *highest_run_time;
>> static struct schedtm_data *lowest_run_time;
>> static struct schedtm_data *highest_latency;
>> static struct schedtm_data *lowest_latency;
>>
>> static unsigned int normalnr = 1;
>> static unsigned int period = 10;
>>
>> module_param(normalnr, uint, 0444);
>> MODULE_PARM_DESC(normalnr, "Number of normal threads");
>>
>> module_param(period, uint, 0444);
>> MODULE_PARM_DESC(period, "Running period in jiffies");
>>
>> void schedtm_sched_in(struct preempt_notifier *pn, int cpu)
>> {
>> struct schedtm_data *data = container_of(pn, struct schedtm_data,
>> pn);
>> WARN_ON(data->t != current);
>>
>> data->run_time_tmp = ktime_to_us(ktime_get());
>> data->latency += ktime_to_us(ktime_get()) - data->latency_tmp;
>> data->latency_count++;
>> data->run_count[cpu]++;
>> }
>>
>> void schedtm_sched_out(struct preempt_notifier *pn, struct task_struct
>> *next)
>> {
>> int cpu = smp_processor_id();
>> struct schedtm_data *data = container_of(pn, struct schedtm_data,
>> pn);
>> WARN_ON(data->t != current);
>>
>> data->run_time += ktime_to_us(ktime_get()) - data->run_time_tmp;
>> data->latency_tmp = ktime_to_us(ktime_get());
>> data->preempt_count[cpu]++;
>> }
>>
>> struct preempt_ops schedtm_po = {
>> .sched_in = schedtm_sched_in,
>> .sched_out = schedtm_sched_out,
>> };
>>
>> static int schedtm_thread_fn(void *arg)
>> {
>> struct schedtm_data *data = (struct schedtm_data *)arg;
>> unsigned long checkp = jiffies;
>>
>> data->start_time = ktime_to_us(ktime_get());
>> data->run_time_tmp = ktime_to_us(ktime_get());
>> data->pn.ops = &schedtm_po;
>> preempt_notifier_register(&data->pn);
>>
>> do {
>> WARN_ON(current != data->t);
>> if (jiffies - checkp > period) {
>> checkp = jiffies;
>> schedule();
>> }
>> } while (!kthread_should_stop());
>>
>> preempt_notifier_unregister(&data->pn);
>> data->total_time = ktime_to_us(ktime_get()) - data->start_time;
>> data->wait_time = data->total_time - data->run_time;
>> data->latency = data->latency / (data->latency_count + 1);
>>
>> return 0;
>> }
>>
>> static void schedtm_report_cpu(struct schedtm_data *data)
>> {
>> int cpu = 0;
>>
>> pr_schedtm("\tcpu count:\tcpu\trun\tpreempt\n");
>>
>> for_each_online_cpu(cpu) {
>> pr_schedtm("\t\t\t%d\t%ld\t%ld\n", cpu,
>> data->run_count[cpu],
>> data->preempt_count[cpu]);
>> }
>> }
>>
>> static void schedtm_print_time(unsigned long us)
>> {
>> /*
>> unsigned long s = us / 1000000;
>> unsigned long ms = (us - s * 1000000) / 1000;
>> us = us - s * 1000000 - ms * 1000;
>> pr_schedtm("\t\t\t%lds%ldms%ldus\n", s, ms, us);
>> */
>> }
>>
>> static void schedtm_report_run(struct schedtm_data *data)
>> {
>> unsigned long rp = (10000 * data->run_time) / (data->total_time + 1);
>> unsigned long rpa = rp / 100;
>> unsigned long rpb = rp - rpa * 100;
>>
>> pr_schedtm("\ttotal time(us):\t%ld\n", data->total_time);
>> schedtm_print_time(data->total_time);
>>
>> pr_schedtm("\trun time(us):\t%ld(%ld.%ld%%)\n",
>> data->run_time, rpa, rpb);
>> schedtm_print_time(data->run_time);
>>
>> pr_schedtm("\twait time(us):\t%ld\n", data->wait_time);
>> schedtm_print_time(data->wait_time);
>> }
>>
>> static void schedtm_report_latency(struct schedtm_data *data)
>> {
>> pr_schedtm("\tlatency(us):\t%ld\n", data->latency);
>> schedtm_print_time(data->latency);
>> }
>>
>> static void schedtm_report_all(void)
>> {
>> /*
>> struct schedtm_data *data = NULL;
>> list_for_each_entry(data, &schedtm_data_list, list) {
>> pr_schedtm("%s\n", data->t->comm);
>>
>> schedtm_report_cpu(data);
>> schedtm_report_run(data);
>> schedtm_report_latency(data);
>> }
>> */
>> }
>>
>> static void schedtm_summary_calculate(struct schedtm_data *summary)
>> {
>> int cpu = 0;
>> struct schedtm_data *data = NULL;
>>
>> list_for_each_entry(data, &schedtm_data_list, list) {
>> for_each_online_cpu(cpu) {
>> summary->run_count[cpu] += data->run_count[cpu];
>> summary->preempt_count[cpu] += data->preempt_count[cpu];
>> }
>>
>> summary->total_time += data->total_time;
>> summary->run_time += data->run_time;
>> summary->wait_time += data->wait_time;
>> summary->latency += data->latency;
>>
>> if (!highest_run_time ||
>> highest_run_time->run_time < data->run_time)
>> highest_run_time = data;
>> if (!lowest_run_time ||
>> lowest_run_time->run_time > data->run_time)
>> lowest_run_time = data;
>> if (!highest_latency ||
>> highest_latency->latency < data->latency)
>> highest_latency = data;
>> if (!lowest_latency ||
>> lowest_latency->latency > data->latency)
>> lowest_latency = data;
>> }
>>
>> summary->total_time /= normalnr;
>> summary->run_time /= normalnr;
>> summary->wait_time /= normalnr;
>> summary->latency /= normalnr;
>> }
>>
>> static void schedtm_report_special(struct schedtm_data *summary)
>> {
>> pr_schedtm("%s got highest run time %ld(+%ld%%)\n",
>> highest_run_time->t->comm,
>> highest_run_time->run_time,
>> (100 * (highest_run_time->run_time - summary->run_time))
>> / (summary->run_time + 1));
>>
>> pr_schedtm("%s got lowest run time %ld(-%ld%%)\n",
>> lowest_run_time->t->comm,
>> lowest_run_time->run_time,
>> (100 * (summary->run_time - lowest_run_time->run_time))
>> / (summary->run_time + 1));
>>
>> pr_schedtm("%s got highest latency %ld(+%ld%%)\n",
>> highest_latency->t->comm,
>> highest_latency->latency,
>> (100 * (highest_latency->latency - summary->latency))
>> / (summary->latency + 1));
>>
>> pr_schedtm("%s got lowest latency %ld(-%ld%%)\n",
>> lowest_latency->t->comm,
>> lowest_latency->latency,
>> (100 * (summary->latency - lowest_latency->latency))
>> / (summary->latency + 1));
>> }
>>
>> static void schedtm_summary_all(void)
>> {
>> struct schedtm_data *summary =
>> kzalloc(sizeof(struct schedtm_data), GFP_KERNEL);
>> if (!summary) {
>> pr_schedtm("failed to summary\n");
>> return;
>> }
>>
>> schedtm_summary_calculate(summary);
>>
>> pr_schedtm("summary\n");
>>
>> schedtm_report_cpu(summary);
>> schedtm_report_run(summary);
>> schedtm_report_latency(summary);
>> schedtm_report_special(summary);
>>
>> kfree(summary);
>> }
>>
>> static int schedtm_succeed(void)
>> {
>> int threadnr = 0;
>> struct schedtm_data *data = NULL;
>>
>> list_for_each_entry(data, &schedtm_data_list, list) {
>> kthread_stop(data->t);
>> threadnr++;
>> }
>>
>> return threadnr == normalnr;
>> }
>>
>> static void schedtm_exit(void)
>> {
>> struct schedtm_data *data = NULL;
>>
>> if (schedtm_succeed()) {
>> schedtm_report_all();
>> schedtm_summary_all();
>> }
>>
>> list_for_each_entry(data, &schedtm_data_list, list)
>> kfree(data);
>>
>> pr_schedtm("scheduler test module off\n");
>> }
>>
>> static int __init schedtm_check_param(void)
>> {
>> if (normalnr == 0) {
>> pr_schedtm("illegal normalnr %d\n", normalnr);
>> return 0;
>> }
>>
>> if (normalnr >= nr_cpu_ids * 100) {
>> pr_schedtm("too big normalnr %d for %d cpu system\n",
>> normalnr, nr_cpu_ids);
>> return 0;
>> }
>>
>> #ifndef CONFIG_PREEMPT
>> if (period >= 100) {
>> pr_schedtm("too big period %d for none kernel-preempt system\n",
>> period);
>> return 0;
>> }
>> #endif
>>
>> return 1;
>> }
>>
>> static int __init schedtm_init(void)
>> {
>> int threadnr = normalnr;
>> struct schedtm_data *data = NULL;
>>
>> if (!schedtm_check_param())
>> return -EAGAIN;
>>
>> pr_schedtm("scheduler test module on\n");
>>
>> while (threadnr--) {
>> data = kzalloc(sizeof(struct schedtm_data), GFP_KERNEL);
>> if (!data)
>> goto failed;
>>
>> data->t = kthread_create(schedtm_thread_fn, data,
>> "stmt%d", threadnr);
>> if (!data->t) {
>> kfree(data);
>> goto failed;
>> }
>>
>> list_add_tail(&data->list, &schedtm_data_list);
>> }
>>
>> list_for_each_entry(data, &schedtm_data_list, list)
>> wake_up_process(data->t);
>>
>> return 0;
>>
>> failed:
>> list_for_each_entry(data, &schedtm_data_list, list) {
>> kthread_stop(data->t);
>> kfree(data);
>> }
>>
>> return -ENOMEM;
>> }
>>
>> module_init(schedtm_init);
>> module_exit(schedtm_exit);
>>
>