add a template.json file that can be used by tune_json.py to create use cases
with various type a load
Signed-off-by: Vincent Guittot <vincent.guittot(a)linaro.org>
---
doc/examples/template.json | 28 ++++++++++++++++++++++++++++
1 file changed, 28 insertions(+)
create mode 100644 doc/examples/template.json
diff --git a/doc/examples/template.json b/doc/examples/template.json
new file mode 100644
index 0000000..5ed3215
--- /dev/null
+++ b/doc/examples/template.json
@@ -0,0 +1,28 @@
+{
+ /*
+ * Simple use case which creates 10% load
+ * for 6 seconds.
+ * A "sleep" : 0 has been added so the file can be used by tune_json.py to
+ * use a sleep event instead of the timer. In this latter case, you need
+ * to set the timer's period to 0
+ */
+ "tasks" : {
+ "thread0" : {
+ "instance" : 1,
+ "loop" : -1,
+ "run" : 10000,
+ "sleep" : 0,
+ "timer" : { "ref" : "unique", "period" : 100000 }
+ }
+ },
+ "global" : {
+ "duration" : 6,
+ "calibration" : "CPU0",
+ "default_policy" : "SCHED_OTHER",
+ "pi_enabled" : false,
+ "lock_pages" : false,
+ "logdir" : "./",
+ "log_basename" : "rt-app2",
+ "gnuplot" : true
+ }
+}
--
1.9.1
Calibration sequence fails to get the right ns per loop value when the
calibration is done on the A57 cluster of the mt8173evb.
A new calibration has been added that adds a sleep period between each each
calibration loop. This idle phase enable rt-app to get the to get the right ns
per loop value for such kind of HW.
The calibration sequence finally uses the lowest value of the calibration
methods which match with the highest compute capacity
Reported-by: Koan-Sin Tan <freedom.tan(a)linaro.org>
Signed-off-by: Vincent Guittot <vincent.guittot(a)linaro.org>
---
src/rt-app.c | 80 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++----
1 file changed, 75 insertions(+), 5 deletions(-)
diff --git a/src/rt-app.c b/src/rt-app.c
index 97aee0f..37e9892 100644
--- a/src/rt-app.c
+++ b/src/rt-app.c
@@ -62,12 +62,14 @@ void waste_cpu_cycles(int load_loops)
}
/*
-* calibrate_cpu_cycles()
-* collects the time that waste_cycles runs.
+* calibrate_cpu_cycles_1()
+* 1st method to calibrate the ns per loop value
+* We alternate idle period and run period in order to not trig some hw
+* protection mechanism like thermal mitgation
*/
-int calibrate_cpu_cycles(int clock)
+int calibrate_cpu_cycles_1(int clock)
{
- struct timespec start, stop;
+ struct timespec start, stop, sleep;
int max_load_loop = 10000;
unsigned int diff;
int nsec_per_loop, avg_per_loop = 0;
@@ -75,6 +77,10 @@ int calibrate_cpu_cycles(int clock)
while (cal_trial) {
cal_trial--;
+ sleep.tv_sec = 1;
+ sleep.tv_nsec = 0;
+
+ clock_nanosleep(CLOCK_MONOTONIC, 0, &sleep, NULL);
clock_gettime(clock, &start);
waste_cpu_cycles(max_load_loop);
@@ -100,6 +106,69 @@ int calibrate_cpu_cycles(int clock)
return 0;
}
+/*
+* calibrate_cpu_cycles_2()
+* 2nd method to calibrate the ns per loop value
+* We continously runs something to ensure that CPU is set to max freq by the
+* governor
+*/
+int calibrate_cpu_cycles_2(int clock)
+{
+ struct timespec start, stop, sleep;
+ int max_load_loop = 10000;
+ unsigned int diff;
+ int nsec_per_loop, avg_per_loop = 0;
+ int ret, cal_trial = 1000;
+
+ while (cal_trial) {
+ cal_trial--;
+
+ clock_gettime(clock, &start);
+ waste_cpu_cycles(max_load_loop);
+ clock_gettime(clock, &stop);
+
+ diff = (int)timespec_sub_to_ns(&stop, &start);
+ nsec_per_loop = diff / max_load_loop;
+ avg_per_loop = (avg_per_loop + nsec_per_loop) >> 1;
+
+ /* collect a critical mass of samples.*/
+ if ((abs(nsec_per_loop - avg_per_loop) * 50) < avg_per_loop)
+ return avg_per_loop;
+
+ /*
+ * use several loop duration in order to be sure to not
+ * fall into a specific platform loop duration
+ *(like the cpufreq period)
+ */
+ /*randomize the number of loops and recheck 1000 times*/
+ max_load_loop += 33333;
+ max_load_loop %= 1000000;
+ }
+ return 0;
+}
+
+/*
+* calibrate_cpu_cycles()
+* Use several methods to calibrate the ns per loop and get the min value which
+* correspond to the highest achievable compute capacity.
+*/
+int calibrate_cpu_cycles(int clock)
+{
+ int calib1, calib2;
+
+ /* Run 1st method */
+ calib1 = calibrate_cpu_cycles_1(clock);
+
+ /* Run 2nd method */
+ calib2 = calibrate_cpu_cycles_2(clock);
+
+ if (calib1 < calib2)
+ return calib1;
+ else
+ return calib2;
+
+}
+
static inline loadwait(unsigned long exec)
{
unsigned long load_count;
@@ -531,13 +600,14 @@ int main(int argc, char* argv[])
/* Needs to calibrate 'calib_cpu' core */
if (opts.calib_ns_per_loop == 0) {
+ log_notice("Calibrate ns per loop");
cpu_set_t calib_set;
CPU_ZERO(&calib_set);
CPU_SET(opts.calib_cpu, &calib_set);
sched_getaffinity(0, sizeof(cpu_set_t), &orig_set);
sched_setaffinity(0, sizeof(cpu_set_t), &calib_set);
- p_load = calibrate_cpu_cycles(CLOCK_THREAD_CPUTIME_ID);
+ p_load = calibrate_cpu_cycles(CLOCK_MONOTONIC);
sched_setaffinity(0, sizeof(cpu_set_t), &orig_set);
log_notice("pLoad = %dns : calib_cpu %d", p_load, opts.calib_cpu);
} else {
--
1.9.1
These series includes fixes/changes that have been done. These patches
mainly fixed some typos or moves code but should not change the
behavior of rt-app.
Those patches are also found in the branch:
https://git.linaro.org/people/picheng.chen/rt-app.git fix
Chris Muller (1):
Update thread name
Ivan T. Ivanov (1):
fixup json-c dependencies
Vincent Guittot (12):
Minor clean up
fix deadline print format
consolidate trace and debug point
update .gitignore
fix inconsistency in delay unit
fix cpu affinity string info
deadline: set deadline field to deadline parameter
reorder the start sequence of threads
cleanup of doc directory
remove useless json_object_put
remove conditional compilation of json Remove useless --with-json
rt-app: remove use of deprecated json interface
.gitignore | 8 ++
README.in | 7 +-
autogen.sh | 2 +-
configure.ac | 17 ++--
doc/taskset.yml | 53 -----------
src/Makefile.am | 8 +-
src/rt-app.c | 227 ++++++++++++++++++++++++++--------------------
src/rt-app_args.c | 77 ++++++++--------
src/rt-app_parse_config.c | 51 +++++++----
src/rt-app_parse_config.h | 2 +-
src/rt-app_utils.c | 3 +
src/rt-app_utils.h | 4 +-
12 files changed, 228 insertions(+), 231 deletions(-)
delete mode 100644 doc/taskset.yml
--
1.9.1