Bad multicore performance

[Mirzet Kadic]

I ran a couple of cpu and video recording tests with some onboard tools.

I realized that almost everything is runing on one core. In best case are two cores used.

Here is a presentation:
https://docs.google.com/presentation/d/1jyxGkxqJLp5-WOpmaCva8Mkwc0ivJp0ZtC91gCmrMQI/edit

Why that?

You can download all files fom the presentation here: https://docs.google.com/open?id=0B4yEtafZsxezaUpURjRyTUxFMG8

[Matt Fischer (mfisch)]

Can you provide some comparison between the N7 and another device? I'd like to confirm that this is or is not platform specific.

[Brendan Donegan (brendan-donegan)]

Hi,

Can you clarify that what you mean is that all the cores are not being used effectively? This may be a bit too wide-ranging a bug, but I'll let someone else make that call.

[Chris Wayne (cwayne)]

I'm not able to reproduce this. Running htop while running different programs, I'm seeing all cores being used. Please provide more information on how to reproduce.

[Mirzet Kadic (caracal-enl)]

I am not a hacker or dev. I don't no if this is a problem of the kernel or the apps. Could it be that the benchmarks aren't optimized for multicore?. I compared the Tablet with my Desktop with Core 2 Duo and I used PowerTOP. Here are the new results:
https://docs.google.com/presentation/d/12ZXc5dhkZbDxElL3frRrltyYgDBtQt3nJnMaGjkandk/edit

However my feeling in general is that the response of the tablet is very low. Recording a video and and voice using RecordMyDesktop or Kazam was painfully slow on this monster machine with 4 CPU Cores and 8 GPU Cores. The CUP Benchmarks were faster on my Desktop than N7.

How to reproduce:
Run the same apps as I used and see if you get the same results. :-|

[Anders Østerholt (diebels)]

Caracal,
can you run multiple instances of hardinfo simultaneously to see if more cores gets used?

By the way, Tegra3 is not that powerful, more like Intel Atom level of performance.

[Chris Wayne (cwayne)]

I ran sysbench --test=cpu --num-threads=4 on this system and on my i5 laptop. The i5 shows the following:

Test execution summary:
    total time: 3.3385s
    total number of events: 10000
    total time taken by event execution: 13.3466
    per-request statistics:
         min: 1.21ms
         avg: 1.33ms
         max: 15.67ms
         approx. 95 percentile: 1.38ms

Threads fairness:
    events (avg/stddev): 2500.0000/38.70
    execution time (avg/stddev): 3.3367/0.00

While the nexus 7 results are attached. Notice the enormous difference (total time: i5 is 3.3, nexus7 is 48.8)

Due to these results, marking back as confirmed.

[Mirzet Kadic (caracal-enl)]

I'm really relieved...
Thank you Chris.
Cannot provide you with future test results for a couple of days due to 4.2 :-|

[Oliver Grawert (ogra)]

We switched to the interactive cpufreq governor a while ago which should act a lot more fine grained ... could someone check if the issue still exists with a recent image