trip-points throttling?

Post Reply
blu
Posts: 74
Joined: Wed Mar 08, 2017 11:30 pm
languages_spoken: english
ODROIDs: XU4, N2
Has thanked: 3 times
Been thanked: 19 times
Contact:

trip-points throttling?

Unread post by blu » Sat Apr 13, 2019 5:53 pm

Heya,

I've been benchmarking some code on a XU4 (active cooling version) with Memeka's excellent Stretchy Pants image, kernel 4.9.11, and I've noticed a what I suspect is a thermal-throttling-related performance oddity:

1. setting the big core's governor to performance (2GHz)
2. pinning a computational single-thread load to the big cores (taskset 0xf0)
3. performance (i.e. time) oscillates between 20s and 22s for the same workload

Checking the thermal zones, temp at idle for zone0 is approx 45-47C:

Code: Select all

$ cat /sys/class/thermal/thermal_zone0/temp
45000
At the same time, the trip points are set as:

Code: Select all

$ for i in {0..3}; do cat /sys/class/thermal/thermal_zone0/trip_point_${i}_temp ; done
50000
60000
70000
120000
Monitoring temp-freq on the big cluster as:

$ while (true); do cat /sys/class/thermal/thermal_zone0/temp ; cat /sys/devices/system/cpu/cpufreq/policy4/scaling_cur_freq ; sleep 1; done

Shows that for the duration of the benchmark temp rises to mid 60's C, and the fan kicks in at 50C, but the freq stays stable at 2GHz. And yet, benchmark times oscillate by 10%. The only IO the benchmark does is some printf, which is redirected to /dev/null. Any ideas what might be the cause of oscillations?

joy
Posts: 881
Joined: Fri Oct 02, 2015 1:44 pm
languages_spoken: english
ODROIDs: ODROID-C1+, XU4, X
Has thanked: 11 times
Been thanked: 47 times
Contact:

Re: trip-points throttling?

Unread post by joy » Fri Apr 19, 2019 12:13 pm

blu wrote:
Sat Apr 13, 2019 5:53 pm
And yet, benchmark times oscillate by 10%. The only IO the benchmark does is some printf, which is redirected to /dev/null. Any ideas what might be the cause of oscillations?
Do you mean your benchmark score is changed even in case of temperature 60C case?

blu
Posts: 74
Joined: Wed Mar 08, 2017 11:30 pm
languages_spoken: english
ODROIDs: XU4, N2
Has thanked: 3 times
Been thanked: 19 times
Contact:

Re: trip-points throttling?

Unread post by blu » Sat Apr 20, 2019 8:31 pm

joy wrote:
Fri Apr 19, 2019 12:13 pm
Do you mean your benchmark score is changed even in case of temperature 60C case?
I've noticed that times remain nominal (i.e. low) if the temp does not pass 50C (fan not kicking in). That happens rarely though, as it takes rather low ambient temps.

Tomorrow I will try to rule out the last non-throttling possibility for the oscillations: variations in the mmap impact of the bench caused by heap access patterns.

Post Reply

Return to “General Topics”

Who is online

Users browsing this forum: No registered users and 0 guests