Hi,
On 10/30/2014 05:04 PM, Cyril Hrubis wrote:
> Hi!
>> #define _GNU_SOURCE
>> @@ -30,9 +33,11 @@
>> #include <stdlib.h>
>> #include <sched.h>
>> #include <sys/time.h>
>> +#include <sys/resource.h>
>> #include <pthread.h>
>> #include <sys/types.h>
>> #include <sys/stat.h>
>> +#include <unistd.h>
>> #include <fcntl.h>
>>
>> #include "test.h"
>> @@ -54,6 +59,8 @@ const char setspeed[] = SYSFS_CPU_DIR
>> "cpu0/cpufreq/scaling_setspeed";
>> const char curspeed[] = SYSFS_CPU_DIR "cpu0/cpufreq/cpuinfo_cur_freq";
>> const char maxspeed[] = SYSFS_CPU_DIR "cpu0/cpufreq/scaling_max_freq";
>>
>> +const int units = 1000000; /* Mhz */
>> +
>> static void cleanup(void)
>> {
>> SAFE_FILE_PRINTF(NULL, boost, "%d", boost_value);
>> @@ -110,26 +117,34 @@ void *thread_fn(void *val)
>> {
>> struct timeval tv_start;
>> struct timeval tv_end;
>> - int i, res = 0;
>> - intptr_t timeout = (intptr_t) val;
>> + int i;
>> + unsigned long long res = 0;
> The res should be marked as volatile to avoid misoptimizations.
>
>> + intptr_t timeout = (intptr_t) val, ret;
>> +
>> + struct sched_param params;
>> + params.sched_priority = sched_get_priority_max(SCHED_FIFO);
>> + if (pthread_setschedparam(pthread_self(), SCHED_FIFO, ¶ms)) {
>> + tst_resm(TWARN | TERRNO,
>> + "failed to set FIFO sched with max priority");
>> + }
>>
>> gettimeofday(&tv_start, NULL);
> We should use CLOCK_MONOTONIC_RAW for measuring the time here to avoid
> failures when ntp daemon changes the system time, etc.
>
>> tst_resm(TINFO, "load CPU0 for %ld sec...", timeout);
>>
>> do {
>> - for (i = 1; i < 10000; ++i)
>> - res += i * i;
>> + for (i = 0; i < units; ++i)
>> + ++res;
> Hmm, are you sure that the number stored in res divided by units and
> timeout really yields the CPU speed?
It's rough calculation, can be close to CPU speed (if certain conditions
apply), but for comparison purposes it may fit.
> I guess that it may, but that would also depend on what compiler does
> with the code and how microcode is implemented in the target CPU, etc.
> Or are you 100% sure that res/units/timeout will yield cpu speed in Mhz
> in all cases?
No.
> If I'm reading the patch right, what it now does it two measurements to
> see if things are faster with boost, that is fine but we should rather
> stop calling the number speed and stop pretending that it's cpu speed in
> Mhz.
You are right, that's why I changed cpu freq to 'avg counter frequency'.
Thanks,
Alexey
------------------------------------------------------------------------------
_______________________________________________
Ltp-list mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/ltp-list
