Hi, I believe that hw.setperf (as well as programs using it, such as apmd -C) change the CPU frequency of only single (arbitrary) cores, leading to higher temperatures than necessary for idle systems. It would be desirable if all cores were affected. (correct me if I'm wrong!)
It doesn't seem to be easy to hack this into e.g. arch/amd64/amd64/est.c since there's no way to set a CPU affinity mask (only a P_CPUPEG for the scheduler, which is probably of limited use in this case). Can anyone who knows the kernel code well comment / suggest a course of action? A user-space implementation seems out of the question too with no user-facing methods available to set CPU affinity. Practical implications: a Supermicro A1SAi based server sitting on my desk gets quite warm at 61°C when idle and supposedly clocked down to half the nominal CPU frequency at hw.setperf=0 (Linux: 43°C). Regards, Marinos -- Marinos Yannikos [email protected]
