Stress Testing Programs
The Cortex-A53 CPU, used in the Raspberry Pi 3, is known to be subject to overheating. Assuming correct software implementation, the first noticeable effect is that, as the temperature increases beyond a critical point, the CPU MHz is throttled. At normal room temperatures, this might only occur when all CPU cores are executing at higher speeds, with a possible contribution from graphics activity. When considered important, special cooling arrangements might be needed, where these stress tests will be of use to evaluate different arrangements. For this series of procedures, the RPI 3 board was “out of case”, where recorded temperatures are often shown to be lower than those obtained using a standard plastic enclosure.
A main consideration for stress testing is that programs have parameters to run for defined durations but with short term reports on progress, including performance and, in this case, CPU temperature and clock MHz. These details should also be saved in constantly updated log files. Then, there will be some evidence, if the system crashes.
In this case, multiple programs are run using a different terminal window for each, normally with 15 minutes test duration specified. One of these measures CPU temperature MHz at specified intervals, where vcgencmd function has to be installed (as used by Raspbian). Two of the programs are benchmarks, already reported on, but with alternative run time parameters, and two are new ones, now with programs, source code and detailed results included in:
http://www.roylongbottom.org.uk/Rpi3-64 ... rks.tar.gz
http://www.roylongbottom.org.uk/Raspber ... hmarks.htm
The oldest uses linverloopsPi64, Livermore Loops Benchmark, that has 24 different test kernels, repeated three time with different (cache) memory demands. This was known to produce wrong numeric answers on an overclocked PC. For reliability testing, a parameter specifies a standard time for each of the loops. Results are displayed during the tests, but performance reported and logged is at the end. The main benefit is a continuously changing processing profile.
The other existing benchmark is videogl64 via OpenGL. This has six tests procedures, where one is chosen besides the number of passes and duration of each. The window width and height can be specified, allowing visible screen space where other terminal windows can be displayed.
The new tests have a run time parameter to specify the amount of cache or memory space to use, and carry out high speed integer and floating point calculations via stressintPi64 and burninfpuPi64.
A summary of test results is;
Integer Arithmetic Stress Test - comprising four runs of stressintPi64 using 40 KB of data, aimed at all using L2 cache, with 12 tests each running for 80 seconds. Performance on all cores was essentially the same, with CPU throttling starting after 30 seconds, eventually reducing CPU MHz by nearly 32%, with maximum recorded sample CPU temperature of 84.4 °C. Compared with stand alone results, CPU performance was degraded to a greater extent due to MP overheads.
Floating Point Arithmetic Stress Test - having four burninfpuPi64 test procedures, using L2 cache with 8 operations per data word. Again performance was effectively constant from all cores, with maximum total throughput of 13.7 GFLOPS, reducing by nearly 4 GFLOPS due to CPU throttling down to 843 MHz, again with a maximum temperature of 84.4 °C.
Livermore Loops Stress Test - This uses four copies of the Livermore Loops Benchmark. Overall MFLOPS speeds are shown to be significantly degraded, but RPiHeatMHz64 results demonstrate inconsistent effects of different arithmetic functions. Maximum temperature recorded was 84.9 °C with a CPU MHz of 744.
Integer and OpenGL Stress Tests - The most complicated OpenGL kitchen test was used, along with three Integer Stress Tests., this time using L1 cache based data. The same procedures were used with CPU MHz settings of On-demand and Performance, where results are shown to be virtually the same. The first summary of speeds and temperatures below is with the Performance setting. Then, OpenGL FPS and integer MB/second reduced to around 60% of initial speeds, with many temperatures of 84.9 °C recorded, when CPU MHz temporarily dropped to half speed at 600 MHz. The tests were repeated with the system in a FLIRC case, where the whole aluminium case becomes the heatsink. The performance was consistently high, but temperatures approached the critical CPU throttling would occur.
Code: Select all
Performance out of case Performance FLIRC case
Total OGL CPU CPU Total OGL CPU CPU
Secs MB/s FPS MHz 'C Secs MB/s FPS MHz 'C
0 1200 55.8 0 1200 44.0
30 13 1107 80.6 30 13 1200 60.1
60 11 910 82.7 60 13 1199 63.4
80 6064 9 850 83.8 80 7116 13 1200 65.0
160 4656 9 744 84.9 160 7041 13 1199 68.8
240 4305 8 600 82.7 240 7072 13 1200 70.9
320 4217 8 600 82.7 320 7075 13 1200 72.0
400 4209 8 738 84.9 400 7095 13 1200 74.1
480 4209 8 600 82.7 480 7081 13 1200 75.8
560 4802 8 738 84.9 560 8067 13 1200 74.7
640 4768 8 722 84.9 640 8092 13 1200 76.8
720 4730 8 743 84.9 720 7989 13 1200 77.4
800 4664 8 823 84.9 800 8050 13 1200 78.4
880 4712 8 719 84.9 880 7984 13 1200 79.5
960 5917 8 938 82.7 960 8344 13 1200 74.1
These are the last of my current benchmarks and test programs for Raspberry Pi 3.
