I found that my "PCI Adapter" overheats when load is above 4.0 (having 4 cores). It does NOT overheat when the load is below 4.0 (but I had an overclocked memory that was causing trouble as I explain at the end), so my focus is on lowering the system load. What overheats is just the k10temp-pci-00c3 PCI Adapter
(according to sensors).
The mistake:
I am trying to track applications that compose that value, and how much each did.
From this question, I found atop
, but reading its output, it is not clear what applications composed the system load... It is like a "simple" column showing SYSLOAD is missing. It seems I could use all that displayed information to calculate something that could show such column, but I just don't know how.
The explanations I found about system load seems too generic (I can be wrong..), so I couldn't understand them enough to produce an algorithm; I think I should mix some data like cpu usage, io usage, mem usage and so get the sysload even if it is somewhat fuzzily guessed...
Precise info on how system Load works:
on this question there is a reference to a pdf explaining it. Load seems not exactly related to specific applications...
Still need a workaround:
Anyway I still would like to know how possible it is to make an application approach about system load, may be there is some way to prioritize the whole system utilization to some specific application, while all the others clogging it, even when they are NOT using much cpu (less than 3%), calm down?
Tests on-going:
I set all processes to nice -n 19
, but the one window/process I am actively using; I saw improvement on the temperature graph with psensor (the graph is a lot less fuzzy); the temperature is holding longer, below the limit, now; all processes seems to be behaving as expected; the system load is still high but I saw it once go til 6.0 and the temperature was still fine; no sudden shutdowns since.. need more test tho...
Finally the culprit was found:
My memory was supposed to accept 2000MHz. Long ago I set it to work at 1600MHz, because the machine would not boot otherwise. Now I set it to automatic, what made it work at 1333MHz. The temperature k10temp-pci-00c3 that was overheating (reaching above 80c) now is consistent around 60c when I run heavy applications!
Also the load average is around 3.5 to 4.0 without causing any hardware problems!
The nice 19 on processes helped postpone the overheating indeed, but wasnt able to totally prevent it!
top
and monitor its output?kill -SIGSTOP ...
on a process that was using 130% cpu (4core max would be 400% I think), the cpu load went from 4.2 to 2.8, and in the list of process using cpu it shows several applications using less than 5% cpu.. Xorg using 15%.. my guess is the load is elsewhere, not something I can guess by reading cpu percentuals. I got about 18xterm
running on a single viewport on compiz, each with its script for monitoring and other tasks; one of them had fast/high text output, I modified it and helped on lowering the load; but the others refresh each 5 seconds;what I do?sensors
it is almost certainly the CPU temperature. On AMD quad-core cpus (among others), the PCI controller has migrated onto the CPU chip from its historically separate "northbridge" chip.