Question about the missing of Graphics Clock Setting

[FuryMartin]

Hi, thanks for building this project, which is a wonderful tool to monitor and optimize GPU energy consumption.

I have a question about clock setting and wish to have a discussion. I carefully checked the code and found Zeus implementing frequency control in zeus/optimizer/perseus/frequency_controller.py. However, it seems that there are only Memory Clock settings.

I am wondering why Graphics Clock setting is missing:

• Does it have no impact on computing energy consumption, and there is no need to control it?
• Or is it because Graphics Clock setting has not been implemented by pynvml? (I have checked pynvml and found nothing about Graphics Clock setting function, so I guess this might be a possible reason?)

[jaywonchung]

Hi @FuryMartin!

The Perseus optimizer controls the GPU primarily in terms of its graphics clock: https://github.com/ml-energy/zeus/blob/b0c65a4e8a2271e34f20c9a40949b01b676ceaa5/zeus/optimizer/perseus/frequency_controller.py#L75

The corresponding NVML API is documented here.

Memory clock can in general be ignored, at least for the GPUs we have dealt with. For instance, A100 supports only one memory clock frequency, and A40 supports two, but one is basically the default and the other is almost zero, putting memory into a pseudo-sleep state. In any case, we're just making it sure memory is running at the highest frequency at all times, and computation time and energy is controlled by setting the graphics clock.

[FuryMartin]

Thanks for your thorough explanation, sorry for having a misunderstanding about the clock control.

I may have been misled by the parameter names minMemClockMHz and maxMemClockMHz in zeus/device/gpu.py, which seems to be setting the Memory Clock:

    @abc.abstractmethod
    def setGpuLockedClocks(
        self, index: int, minMemClockMHz: int, maxMemClockMHz: int
    ) -> None:
        """Lock the GPU clock to a specified range. Units: MHz."""

[jaywonchung]

Oh, you're right. Those parameters have wrong names. Thanks a lot for catching those! Would you mind sending a quick PR to fix the name so that they're aligned with the NVML API? If you don't have the bandwidth, I can quickly fix them.

[FuryMartin]

It's an honor to do it. I will finish it quickly.😊