Why different runtimes for constant exponent and (huge) modulus for mpz_powm()?

[hermann at stamm-wilbrandt.de]

On 2023-07-25 09:02, Niels Möller wrote:
> 
> Curious. I would suspect some enviromental factors (e.g., high 
> processor
> temperature sometimes leading to some frquency throttling, or something
> else on the machine periodically competing for cpu).
> 
Thanks for responding.
That machine runs completely alone.
It just completed 908 rounds of the 11121 rounds.
The pattern changed a bit, once there was an upward phase before going 
back:
https://stamm-wilbrandt.de/en/forum/908.png

Eich of the 908 round on x-axis takes nearly 10 minutes, so plateaus 
seen take a long time. Im am on that machine only short times for 
getting latest data or so.

Because of Ryzen 7000 series burnout risk I applied latest stable 1.24 
bios before that run. 7600X CPU max operating temperature is 95°C, and I 
operated it with PBO 85°C before bios upgrade. But after upgrade I 
cannot set it above 75°C anymore, and maximally measured temperature 
with stress test on all threads showed 76°C. So high temperature cannot 
be it, and not throttling.

I did "perf stat process" before starting the now >6 days run and saw 
that CPU frequency was always in 5.2..5.4GHz range (5.3GHz is nominal 
burst CPU frequency).


Here the current state of the PC:

hermann at 7600x:~/9383761-digit-prime$ uptime
  21:46:24 up 6 days,  4:12,  1 user,  load average: 1.00, 1.00, 1.00
hermann at 7600x:~/9383761-digit-prime$
hermann at 7600x:~/9383761-digit-prime$ ps -ef | grep job
root      629465  629454 99 Jul22 ?        4-04:00:15 ./job 297
hermann  1947666 1942377  0 21:46 pts/0    00:00:00 grep --color=auto 
job
hermann at 7600x:~/9383761-digit-prime$

I had to do a hotfix after 2 days of computation to avoid bug that was 
guaranteed to happen after 75 day, details here:
https://github.com/Hermann-SW/9383761-digit-prime#hotfix

Total non-computing time by hotfix tool was less than 10 seconds only.


> Maybe it would be helpful to monitor performance related measurements
> such as cpu frequency (or cycle counter), temperature, and count of
> cache misses as computation progresses?
> 
AMD is not good at temperature measurement support for Linux, but I 
found an article with kernel driver change and then all will be fine. I 
do not want to risk current >70 day computation though, so on Linux side 
I am blind temperature wise at the moment (all tools get wrong values 
with bad kernel support).

But I just learned that I can attach perf to a running process, and did 
that, two times 5 seconds, and once 15 seconds. Always showing 5.23GHz, 
near nominal 5.3GHz burst frequency.

How do you want me to count cache misses, with perf?
Over longer timeframe I think, multiple of 10min/round?


hermann at 7600x:~/9383761-digit-prime$ ps -ef | grep job
root      629465  629454 99 Jul22 ?        4-04:08:14 ./job 297
hermann  1949425 1942377  0 21:54 pts/0    00:00:00 grep --color=auto 
job
hermann at 7600x:~/9383761-digit-prime$ echo 1 | sudo tee 
/proc/sys/kernel/perf_event_paranoid
[sudo] password for hermann:
1
hermann at 7600x:~/9383761-digit-prime$ sudo perf stat -e cycles,task-clock 
-p 629465 sleep 5

  Performance counter stats for process id '629465':

     26,173,724,427      cycles                    #    5.234 GHz
           5,000.98 msec task-clock                #    1.000 CPUs 
utilized

        5.001249395 seconds time elapsed

hermann at 7600x:~/9383761-digit-prime$ sudo perf stat -e cycles,task-clock 
-p 629465 sleep 5

  Performance counter stats for process id '629465':

     26,156,801,154      cycles                    #    5.231 GHz
           5,000.58 msec task-clock                #    1.000 CPUs 
utilized

        5.000851191 seconds time elapsed

hermann at 7600x:~/9383761-digit-prime$ sudo perf stat -e cycles,task-clock 
-p 629465 sleep 15

  Performance counter stats for process id '629465':

     78,483,835,106      cycles                    #    5.232 GHz
          14,999.67 msec task-clock                #    1.000 CPUs 
utilized

       15.000559563 seconds time elapsed

hermann at 7600x:~/9383761-digit-prime$


Regards,

Hermann.