4 ms·
13% cpu usage at the lowest c-state is a also very different than 13% at an elevated c-states. I've recently spent a lot of time analyzing c-states/p-states and
by mmarks 10y ago
13% cpu usage at the lowest c-state is a also very different than 13% at an elevated c-states. I've recently spent a lot of time analyzing c-states/p-states and the power mgmt modes of the GPU. After learning more about the complexity behind the clocks, bus speeds, etc. underlying each state, whenever I hear someone quote a utilization number of a minor workload, I want to know at what power state.
Not to take away but the author's point, just an aside that utilization numbers can be a lot more complicated when there are dozens of energy states and the utilization might be utilization at a particular state rather than utilization at maximum power
- aarongolliver 10y agoYou mean p-states in your first sentence, right? Anything but c0 represents different levels of 'retiring no instructions' (totally idle). The rest of your comment seems accurate though.
- mmarks 10y agoYou're correct that I just meant a lower power state. The author mentioned blinking cursor, so it reminded me of graphics issues. A more efficient CPU state has the possibility of slowing an app due to CPU-GPU sync points. A blocking CPU in an energy efficient state can reawaken slower from GPU done notifications, so FPS is lower. So both c-state and p-states can affect performance. General point was just utilization may not be utilization at max power. I've worked on problems where utilization was 15% at lower power and it was a problem. But to compare different workloads, it'd be < 1% at max power.
- rasz 10y agoeven the best case scenario is 13% of ~900MHz, 100MHz to blink 20 pixels.
- gpderetta 10y agoThat's actually even worse: it will prevent the cpu to enter deep sleep states and save more power. A while ago I was trying to minimize the power usage of my laptop (yeah, slow day) to maximize my battery time. Armed with powertop, I removed any undesired process until I was left with an otherwise idle emacs (less than 1<% cpu) as the last major source of wakeups. Sure enough, disabling the blinking cursor brought that down to nothing, allowing the cpu to stay into deep sleep state much longer.