Am Mittwoch, den 04.11.2015, 18:39 +0100 schrieb Timo Sigurdsson: > Hi, > > since I have an extra A20 board that doesn't have a specific purpose at the > moment, I thought I do some experiments with it and test performance and > reliability (among other things). > One of my tests involved overclocking the board at stock voltage (1.4V) to > see what the board can do. So I used cpufreq-ljt-stress-test and cpuburn-a7 > as mentioned on the wiki[1]. What surprised me, though, was that the stress > test neither runs for a very long time nor on both cores simultaniously. So, > the test results suggest that my board can do 1104MHz at 1.4V (I didn't try > higher frequencies because I didn't even expect that it would run stable at > 1104 MHz without raising the voltage). > > But I'm wondering - how reliable are these tests actually? I would have > assumed that for the results to be meaningful it would be best to put as much > stress on the CPU as possible and do that over a prolonged period. And if you > have multiple cores, to put them under load simultaniously. It this > assumption wrong? Is such extensive testing neglible in real life? > > Since I didn't trust the quick test, I quickly changed the script to to 600 > iterations instead of 60. But, of course, that doesn't make it run on two > cores. So, while running cpufreq-ljt-stress-test, I also ran cpuburn-a7 in > the background which put both cores under load. The board still passed the > test. > > What kind of tests or setups do you use to determine reliable settings? > > > Thanks, > > Timo > > > [1] http://linux-sunxi.org/Hardware_Reliability_Tests
I did some more testing and can answer the main question myself now: I reran my tests again, this time with frequencies up to 1200MHz@1.4V. The unchanged cpufreq-ljt-stress-test fails at 1200MHz but passes at 1152MHz. If I change the test duration by increasing the iterations to 600 and have cpuburn-a7 running in the background while running the stress test, the 1152MHz setting runs fine for quite a while but eventuelly fails on the second core at a late stage - just a few more iterations until it would pass. At 1104MHz it still passes the test. So I guess it's better to take the test result of the unmodified script with a grain of salt and test with more intensive workloads. I'd still be interested to hear what kind of reliable test setups are used or recommended. Maybe we can add some recommendations to the wiki then. Thanks, Timo -- You received this message because you are subscribed to the Google Groups "linux-sunxi" group. To unsubscribe from this group and stop receiving emails from it, send an email to linux-sunxi+unsubscr...@googlegroups.com. For more options, visit https://groups.google.com/d/optout.
