Fix thermal issues

bronco · Published in 2015-11-30 13:15:53

Edited by bronco at 2015-12-14 04:19

I've written a script that does the following:

Installs sunxi-tools if not available
Converts script.bin to a temporary file
Replaces therein the dvfs table settings with more sane values
Makes a backup script.bin.bak before overwriting script.bin with the new settings

EDIT: In the meantime I adopted the official linux-sunxi dvfs/thermal settings -- compare with the thread on the linux-sunxi mailing list.

The script can be found here: http://kaiser-edv.de/tmp/H9rWPf/fix-thermal-problems.sh

Install it as eg. /usr/local/bin/fix-thermal-problems.sh, make it executable and check stability afterwards using the usual approach: http://linux-sunxi.org/Hardware_Reliability_Tests#CPU

wget -O /usr/local/bin/fix-thermal-problems.sh "http://kaiser-edv.de/tmp/H9rWPf/fix-thermal-problems.sh"
chmod 755 /usr/local/bin/fix-thermal-problems.sh

Copy code

Feedback welcome.

hojnikb · Published in 2015-12-1 04:49:59

Cortex a7 running over 2ghz ? i have a hard time believing this...

besides, this thing just looks like more cores slapped together.... it would be more interesting if it used something like 4 a53 cores with 2 a72 fast ones.

bronco · Published in 2015-11-30 16:49:22

Running 4 threads sysbench, 4 threads cpuburn-a7 in parallel while letting walk cpufreq-ljt-stress-test through all available cpufrequencies. No fan used, just a 0.5$ heatsink:
Temperature when idle 39°C (16°C above ambient temperature), under full load at maximum speed 65°C (42°C above ambient temperature), no stability issues. When a display is connected then the H3 switches on its display engine so temperatures increase by approx. 4°C.

root@OrangePI:~# cpufreq-ljt-stress-test
The cjpeg and djpeg tools from libjpeg-turbo are not found.
Trying to download and compile them.
CPU stress test, which is doing JPEG decoding by libjpeg-turbo
at different cpufreq operating points.
Testing CPU 0
1536 MHz SKIPPED
1440 MHz SKIPPED
1344 MHz SKIPPED
1200 MHz ............................................................ OK
1104 MHz ............................................................ OK
1008 MHz ............................................................ OK
912 MHz ............................................................ OK
816 MHz ............................................................ OK
720 MHz ............................................................ OK
648 MHz ............................................................ OK
600 MHz ............................................................ OK
504 MHz ............................................................ OK
480 MHz ............................................................ OK
408 MHz SKIPPED
312 MHz SKIPPED
240 MHz SKIPPED
120 MHz SKIPPED
60 MHz SKIPPED
Testing CPU 1
1536 MHz SKIPPED
1440 MHz SKIPPED
1344 MHz SKIPPED
1200 MHz ............................................................ OK
1104 MHz ............................................................ OK
1008 MHz ............................................................ OK
912 MHz ............................................................ OK
816 MHz ............................................................ OK
720 MHz ............................................................ OK
648 MHz ............................................................ OK
600 MHz ............................................................ OK
504 MHz ............................................................ OK
480 MHz ............................................................ OK
408 MHz SKIPPED
312 MHz SKIPPED
240 MHz SKIPPED
120 MHz SKIPPED
60 MHz SKIPPED
Testing CPU 2
1536 MHz SKIPPED
1440 MHz SKIPPED
1344 MHz SKIPPED
1200 MHz ............................................................ OK
1104 MHz ............................................................ OK
1008 MHz ............................................................ OK
912 MHz ............................................................ OK
816 MHz ............................................................ OK
720 MHz ............................................................ OK
648 MHz ............................................................ OK
600 MHz ............................................................ OK
504 MHz ............................................................ OK
480 MHz ............................................................ OK
408 MHz SKIPPED
312 MHz SKIPPED
240 MHz SKIPPED
120 MHz SKIPPED
60 MHz SKIPPED
Testing CPU 3
1536 MHz SKIPPED
1440 MHz SKIPPED
1344 MHz SKIPPED
1200 MHz ............................................................ OK
1104 MHz ............................................................ OK
1008 MHz ............................................................ OK
912 MHz ............................................................ OK
816 MHz ............................................................ OK
720 MHz ............................................................ OK
648 MHz ............................................................ OK
600 MHz ............................................................ OK
504 MHz ............................................................ OK
480 MHz ............................................................ OK
408 MHz SKIPPED
312 MHz SKIPPED
240 MHz SKIPPED
120 MHz SKIPPED
60 MHz SKIPPED
Overall result : PASSED

Copy code

hojnikb · Published in 2015-12-1 03:09:39

Great stuff. Should be included in every distro from now on

Btw, how is auto undervolting script coming along ?

bronco · Published in 2015-12-1 03:36:43

Edited by bronco at 2015-12-1 03:45

hojnikb replied at 2015-12-1 03:09
how is auto undervolting script coming along ?

Maybe I drop the whole efforts since measurements have shown that you don't win that much by further decreasing voltages while risking stability issues. These settings seem sane to me for normal use cases: http://linux-sunxi.org/User_talk ... and_USB_peripherals
On the other hand SinoVoip sent me a review (or reward?) sample I'm currently playing with. BPi-M3 with A83T SoC. It was a bit hard to get this thing booting (since SinoVoip provides 4 OS images, one useless -- Android -- and the three Linux images all corrupted).

Thx to loboris I now use his image on the BPi-M3 and started with first tests: Without a heatsink it overheats really fast and throttling occurs at 75°C-80°C. I then applied the same $0.5 heatsink I already put on the H3 and this helped a lot: Left is without heatsink, higher temperatures and lower clockspeeds, on the right while running the benchmarks -- see below:

And this SoC is not an ultra-slow one like the H3 but fast instead. Just a heatsink, no fan and no overclocking and it's already twice as fast as Orange Pis. Seems like that with improved cooling the A83T will dynamically clock up to 2.1 GHz and remains there still without overvolting/overclocking

root@OrangePI:~# sysbench --test=cpu --cpu-max-prime=20000 run --num-threads=8
sysbench 0.4.12: multi-threaded system evaluation benchmark
Running the test with following options:
Number of threads: 8
Doing CPU performance benchmark
Threads started!
Done.
Maximum prime number checked in CPU test: 20000
Test execution summary:
total time: 66.9015s
total number of events: 10000
total time taken by event execution: 535.0102
per-request statistics:
min: 41.86ms
avg: 53.50ms
max: 162.34ms
approx. 95 percentile: 63.01ms
Threads fairness:
events (avg/stddev): 1250.0000/3.32
execution time (avg/stddev): 66.8763/0.02
root@OrangePI:~# 7z
7z 7za
root@OrangePI:~# 7za b
7-Zip (A) 9.20 Copyright (c) 1999-2010 Igor Pavlov 2010-11-18
p7zip Version 9.20 (locale=en_US.UTF-8,Utf16=on,HugeFiles=on,8 CPUs)
RAM size: 2012 MB, # CPU hardware threads: 8
RAM usage: 1701 MB, # Benchmark threads: 8
Dict Compressing | Decompressing
Speed Usage R/U Rating | Speed Usage R/U Rating
KB/s % MIPS MIPS | KB/s % MIPS MIPS
22: 3155 674 455 3069 | 82575 786 947 7447
23: 3060 681 458 3118 | 81850 788 949 7488
24: 2347 585 431 2524 | 77849 782 923 7221
25: 1712 454 430 1955 | 77531 779 935 7290
----------------------------------------------------------------
Avr: 598 443 2666 784 939 7362
Tot: 691 691 5014
root@OrangePI:~# mbw -t0 256
Long uses 4 bytes. Allocating 2*67108864 elements = 536870912 bytes of memory.
Getting down to business... Doing 10 runs per test.
0 Method: MEMCPY Elapsed: 0.51003 MiB: 256.00000 Copy: 501.932 MiB/s
1 Method: MEMCPY Elapsed: 0.52706 MiB: 256.00000 Copy: 485.713 MiB/s
2 Method: MEMCPY Elapsed: 0.51487 MiB: 256.00000 Copy: 497.210 MiB/s
3 Method: MEMCPY Elapsed: 0.50704 MiB: 256.00000 Copy: 504.895 MiB/s
4 Method: MEMCPY Elapsed: 0.50709 MiB: 256.00000 Copy: 504.837 MiB/s
5 Method: MEMCPY Elapsed: 0.50694 MiB: 256.00000 Copy: 504.989 MiB/s
6 Method: MEMCPY Elapsed: 0.50748 MiB: 256.00000 Copy: 504.448 MiB/s
7 Method: MEMCPY Elapsed: 0.50718 MiB: 256.00000 Copy: 504.750 MiB/s
8 Method: MEMCPY Elapsed: 0.50719 MiB: 256.00000 Copy: 504.746 MiB/s
9 Method: MEMCPY Elapsed: 0.50771 MiB: 256.00000 Copy: 504.222 MiB/s
AVG Method: MEMCPY Elapsed: 0.51026 MiB: 256.00000 Copy: 501.705 MiB/s
root@OrangePI:~# mbw -t1 256
Long uses 4 bytes. Allocating 2*67108864 elements = 536870912 bytes of memory.
Getting down to business... Doing 10 runs per test.
0 Method: DUMB Elapsed: 0.57809 MiB: 256.00000 Copy: 442.839 MiB/s
1 Method: DUMB Elapsed: 0.57037 MiB: 256.00000 Copy: 448.831 MiB/s
2 Method: DUMB Elapsed: 0.57013 MiB: 256.00000 Copy: 449.016 MiB/s
3 Method: DUMB Elapsed: 0.56990 MiB: 256.00000 Copy: 449.201 MiB/s
4 Method: DUMB Elapsed: 0.56987 MiB: 256.00000 Copy: 449.224 MiB/s
5 Method: DUMB Elapsed: 0.57000 MiB: 256.00000 Copy: 449.120 MiB/s
6 Method: DUMB Elapsed: 0.56974 MiB: 256.00000 Copy: 449.324 MiB/s
7 Method: DUMB Elapsed: 0.57053 MiB: 256.00000 Copy: 448.705 MiB/s
8 Method: DUMB Elapsed: 0.56992 MiB: 256.00000 Copy: 449.187 MiB/s
9 Method: DUMB Elapsed: 0.57001 MiB: 256.00000 Copy: 449.116 MiB/s
AVG Method: DUMB Elapsed: 0.57086 MiB: 256.00000 Copy: 448.448 MiB/s
root@OrangePI:~# mbw -t2 256
Long uses 4 bytes. Allocating 2*67108864 elements = 536870912 bytes of memory.
Using 262144 bytes as blocks for memcpy block copy test.
Getting down to business... Doing 10 runs per test.
0 Method: MCBLOCK Elapsed: 0.50565 MiB: 256.00000 Copy: 506.275 MiB/s
1 Method: MCBLOCK Elapsed: 0.50830 MiB: 256.00000 Copy: 503.641 MiB/s
2 Method: MCBLOCK Elapsed: 0.49063 MiB: 256.00000 Copy: 521.775 MiB/s
3 Method: MCBLOCK Elapsed: 0.49140 MiB: 256.00000 Copy: 520.961 MiB/s
4 Method: MCBLOCK Elapsed: 0.49059 MiB: 256.00000 Copy: 521.822 MiB/s
5 Method: MCBLOCK Elapsed: 0.49041 MiB: 256.00000 Copy: 522.007 MiB/s
6 Method: MCBLOCK Elapsed: 0.48987 MiB: 256.00000 Copy: 522.583 MiB/s
7 Method: MCBLOCK Elapsed: 0.49092 MiB: 256.00000 Copy: 521.468 MiB/s
8 Method: MCBLOCK Elapsed: 0.49066 MiB: 256.00000 Copy: 521.742 MiB/s
9 Method: MCBLOCK Elapsed: 0.49010 MiB: 256.00000 Copy: 522.348 MiB/s
AVG Method: MCBLOCK Elapsed: 0.49385 MiB: 256.00000 Copy: 518.371 MiB/s

Copy code

john0815 · Published in 2015-12-1 04:49:30

OPI 2 get crazy with this script.

Sorry

bronco · Published in 2015-12-1 05:09:09

hojnikb replied at 2015-12-1 04:49
Cortex a7 running over 2ghz ? i have a hard time believing this...

besides, this thing just looks l ...

I thought the same and am still sceptical. More testing has to be done and then we'll see.

For many use cases (file server for example) CPU performance is close to irrelevant. And if it's about I/O performance the M3 really sucks: pseudo-SATA performance 3.5-10 times slower than compared to an A20 based Banana Pi or Orange Pi: http://www.orangepi.org/orangepi ... 8&fromuid=29411

bronco · Published in 2015-12-1 06:44:29

hojnikb replied at 2015-12-1 04:49
Cortex a7 running over 2ghz ? i have a hard time believing this...

besides, this thing just looks l ...

A few last words to A83T since it affects somewhat H3 (performance expectations)

When being clocked identical then the single core performance of H3 and A83T are identical. I tried it with 4 threads at 1008 MHz and 1200 MHz and they perform on par.

The difference is the process, the H3 being made in a cheaper process whereas the A83T as a 'tablet SoC' is produced in the more energy efficient HPC process. That means drastically less voltage needed and higher clockspeeds possible. At the moment 1.8 GHz are configured as maximum and I just tried to adjust the /sys/devices/virtual/thermal/thermal_zone0/trip_point* values. Normally they start at 65°C and increase up to 105° but I adjusted them all up to 95°C except of the last one.

Result: When configured to run at 1.8 GHz max. the A83T remains at this speed on the first run an finished the sysbench test in 52.5 seconds. The next run led to 4 CPU cores being disabled

While it's interesting to see how performant the A83T is during peak performance periods it makes definitely more sense to configure throttling correctly than to accept dropped CPU cores. With increased thermal_zone0/trip_point* values also idle consumption increased a lot. Fortunately CPU peak performance doesn't matter that much in most situations...

hojnikb · Published in 2015-12-2 02:30:37

I see. That makes sense actually.

Too bad we won't get the undervolt script though

It would be great to squeeze every last bit of efficiency out of these boards

bronco · Published in 2015-12-2 06:18:15

hojnikb replied at 2015-12-2 02:30
I see. That makes sense actually.

Too bad we won't get the undervolt script though

We'll see. I'll focus back on H3 since I'm already done with 'Banana Pi M3'. Its SoC is somewhat limited (I/O bandwidth) and the design decisions made by the manufacturer strange (limit bandwidth even more).

http://forum.armbian.com/index.p ... ew-of-banana-pi-m3/

I'm still waiting for the 2nd OPi PC and camera module, then it gets interesting...