load average based scheduling

[Wayne Moore]

I'm thinking about taking on this

Client

1. Change prefs from "if CPU load > X, suspend works" to "if CPU load is X, use N-X CPUs".

The goal is to run ans much boinc as possible without boinc pushing the load average over threshold. I was thinking along the lines of

load = 0;

load2 = 0;

tick(cpu,time)

{

  usage = cpu/time;

  x = exp(-time/TIME_CONSTANT);

  load = x * load + (1 - x) * usage;

  load2 = x * load2 + (1 - x) * usage * usage;

  sd = sqrt(load2 - load * load);

  if (load > threshold + sd)

    if(boinc_running())

      stop_one_boinc();

  if (load < threshold - 2 * sd)

    if (boinc_waiting())

      start_one_boinc();

}

In theory I know how to get (cpu,time) on ubuntu/debian mac and windows but I don't have a Mac development environment and haven't done windows system programming in decades.

[Wayne Moore]

https://github.com/black98fxstc/load-shedding/blob/main/load.h

seems to work quite nicely on Linux and Windows. I stole the Apple code and can't test it, but it looks reasonable. Don't know about other Unices.

Interestingly it works correctly even in a Docker container, with no special privileges, at least on Linux.

[Wayne Moore]

I have a patched client running now and it works quite well. Keeps the load around 6 (out of 12) where I set it. Running in a container.

[Wayne Moore]

FWIW It now will not start jobs unless the cpu_temp < cool_enough and will shed jobs if cpu_temp > too hot, at least on Linux. I understand from the comments someone has done this on Mac. I does not appear to be possible on Windows without an insane amount of work.