Increasing CPU load with release HARK 3.0

HARK FORUM Increasing CPU load with release HARK 3.0

Tagged: ,

This topic contains 6 replies, has 2 voices, and was last updated by Avatar artemyangurazov 2 months, 1 week ago.

Viewing 7 posts - 1 through 7 (of 7 total)
  • Author
    Posts
  • #698
    Avatar
    artemyangurazov
    Participant

    I installed hark 3.0 and found that the network requires 6 times more CPU resources to work.
    The network is used for localization and separation. ROS modules are also used. Linux is used.
    Hark 2.5 CPU load – 20-40% / 800%
    Hark 3.0 CPU load – 120-150% / 800%
    CPU – Intel® Core ™ i7-6700 CPU @ 3.40GHz × 8
    Tell me, please, what could be the reason for such an increase?

    #699
    Avatar
    c.santiago
    Moderator

    To better assist you, would you mind providing the following information.

    1. Is the increase in CPU load average or was it a peak seen in the initialization of the network file.
    2. The Ubuntu and ros version used.
    3. If possible, the network used when this observation was made.

    Best regards,
    c.santiago

    #700
    Avatar
    artemyangurazov
    Participant

    1. This is the average CPU load when the network is running.
    2. OS – Ubuntu 16.04.5 LTS 64-bit, ROS – kinetic
    3. If by using the network you understand the presence of active sources, then on average the load is the same as in the absence of active sources. When new active sources appear, a small peak of load occurs.

    #701
    Avatar
    c.santiago
    Moderator

    I do apologize for not being more specific.

    The network is what we referrer to as the file created and saved using hark_designer.

    We have conducted a few test and have not seen a major increase on the load on our CPU when using HARK 3.0 installed on Ubuntu 16.04. For the tests, the network used localization and separation nodes, to be more specific LocalizeMUSIC and GHDSS.
    We have however, found that the CPU load was increased significantly(up to 60% in our tests) when installing HARK 3.0 on Ubuntu 18.04.
    The source code is the same for both Ubuntu 16.04 and Ubuntu 18.04 so, the reason for the load increase may lie somewhere else.

    As you mentioned, you are currently using Ubuntu 16.04 so the sharp increase in CPU load should not have effected your environment.

    To get to the bottom of the increase in CPU load you have noted, we would like to see the parameters and nodes used in your network file to see if we can replicate your results.

    Best regards,
    c.santiago

    • This reply was modified 2 months, 1 week ago by Avatar c.santiago.
    • This reply was modified 2 months, 1 week ago by Avatar c.santiago.
    • This reply was modified 2 months, 1 week ago by Avatar c.santiago.
    #708
    Avatar
    artemyangurazov
    Participant

    There may be different versions of the system kernel. I have installed Linux version 4.15.0-34-generic.

    Network is attached.

    First argument is device id, second argument is matrix (Localization and Separation) path.

    Attachments:
    1. default_net.n
    #714
    Avatar
    c.santiago
    Moderator

    Thank you for providing your network file.

    We have found the core issue of the problem.

    HARK 2.5 only supported single thread processing, HARK 3.0 on the other hand now uses blas/openblas which allows multi threading.

    The workload is spread over all the available CPU which caused the load you were experiencing.
    Depending on the kind of process, the threads may not be efficiently used.

    To reproduce the same load you had using HARK 2.5, please limit the process to a single thread.

    This can be done by running the following command.

    export OMP_NUM_THREADS=1
    export OPENBLAS_NUM_THREADS=1

    Best regards,
    c.santiago

    • This reply was modified 2 months, 1 week ago by Avatar c.santiago.
    #718
    Avatar
    artemyangurazov
    Participant

    Thank you for the help and quick answers!
    The problem is solved.

Viewing 7 posts - 1 through 7 (of 7 total)

You must be logged in to reply to this topic.