i7 thermal throttling?

[fivemack]

Could someone with an i7 running recent 64-bit Linux do

grep TRM /proc/interrupts

and say if they see numbers even close to the

TRM: 2383324 2150851658 7962 0 7962 0 2150851643 2383324 Thermal event interrupts

figures I have? (and isn't that odd, to have the threads clearly arranged as 01232310?)

powertop is reporting 2000 thermal event interrupts a second, and I'm wondering if the machine is throttling, since an absolutely trivial test program

typedef unsigned long long u64;

#include <stdio.h>

int main(void)
{
  u64 A = 0;
  u64 B = 300000;
  u64 C = 0;
  B = B*B;
  while (A<B) {A++;C-=A;}
  printf("%lld\n",C);
}

runs (timed with time) in

51 seconds utime (core2/2400 on a whole CPU)
51 seconds utime (core2 nice 19 with 4 other nice 19 jobs running)
64 seconds utime (otherwise-idle i7 claiming 2793MHz in /proc/cpuinfo)
180 seconds utime (i7 running tasks in eight threads)
183 seconds utime (i7 running tasks in the threads that taskset calls 0..5 and 7, this job bound to thread #6 with 'taskset 40 ./a.out')
71 seconds utime (Phenom/2500 running tasks on all 4 cores)
71 seconds utime (otherwise-idle Phenom)

[lfm]

Originally Posted by fivemack

Could someone with an i7 running recent 64-bit Linux do

grep TRM /proc/interrupts

and say if they see numbers even close to the

TRM: 2383324 2150851658 7962 0 7962 0 2150851643 2383324 Thermal event interrupts

figures I have? (and isn't that odd, to have the threads clearly arranged as 01232310?)

powertop is reporting 2000 thermal event interrupts a second, and I'm wondering if the machine is throttling, since an absolutely trivial test program

typedef unsigned long long u64;

#include <stdio.h>

int main(void)
{
  u64 A = 0;
  u64 B = 300000;
  u64 C = 0;
  B = B*B;
  while (A<B) {A++;C-=A;}
  printf("%lld\n",C);
}

runs (timed with time) in

51 seconds utime (core2/2400 on a whole CPU)
51 seconds utime (core2 nice 19 with 4 other nice 19 jobs running)
64 seconds utime (otherwise-idle i7 claiming 2793MHz in /proc/cpuinfo)
180 seconds utime (i7 running tasks in eight threads)
183 seconds utime (i7 running tasks in the threads that taskset calls 0..5 and 7, this job bound to thread #6 with 'taskset 40 ./a.out')
71 seconds utime (Phenom/2500 running tasks on all 4 cores)
71 seconds utime (otherwise-idle Phenom)

this code is VERY heavily dependent on the compiler optimizations. I expect slight differences in the compiler versions and/or the optimization options used will account for much of the time differences.

[xorbe]

$ grep TRM /proc/interrupts
TRM: 0 0 0 0 0 0 0 0 Thermal event interrupts

$ uptime
02:53:46 up 4:51, 4 users, load average: 5.51, 4.72, 4.84

i7 @ 3.66GHz 8 threads of P-1, 2 VirtualBoxes, terminals, Firefox, etc

I have CPU-TM (thermal management) disabled in the BIOS (Asus P6T Deluxe), because I saw throttling when compiling the kernel when I first started testing. It's a known thing, as far as I know.

[IronBits]

I always disable thermal management in the BIOS before installing any OS for a 24/7 cruncher to prevent throttling.

[xorbe]

To the OP though -- if you disable CPU-TM, make sure that your cooling solution is good and that you really aren't cooking your cpu. Thoroughly check and test!

[xorbe]

Handy command:

# cat /sys/devices/system/cpu/cpu*/thermal_throttle/count
0
0
0
0
0
0
0
0
#

[fivemack]

Originally Posted by IronBits

I always disable thermal management in the BIOS before installing any OS for a 24/7 cruncher to prevent throttling.

Do you also tape over the little spanner light that suggests your car needs servicing, to prevent engine failure? Throttling's a sign that something elsewhere is wrong; my real issue is how to prevent cpufreq from automatically overclocking my i7 to the point that it throttles. I've written to the author of cpufreq without useful result.

[VJS]

Well since your CPU is giving you the little check engine light perhaps you should take it in for service or exchange???

The check engine light is for idiots most of the time the mechanic will simply reset the light and charge you 50 bucks, if you know what your doing it can be ignored.

IB knows his CPU's as well as thermal management. The point is if you believe intel follow them blindly then let your CPU throttle. If you can monitor your own temperature and optimize your system turn off thermal management and get a good heatsink.

[cheesehead]

Some folks are better at playing with fire, or sword swallowing, than others.

[alpertron]

A cheap approach is (if you have no kids at home) to leave the left side of the case open. I was able to reduce from 79 Celsius degrees to 68 Celsius degrees the temperature of my old Duron 1300 MHz in this way on a very hot summer (up to 35 Celsius degrees in the room). Now I have a Core 2 Quad that runs a lot cooler.

[xorbe]

Originally Posted by fivemack

Do you also tape over the little spanner light that suggests your car needs servicing, to prevent engine failure? Throttling's a sign that something elsewhere is wrong;

No, on i7 at least, part of the throttle function is current limiting, so to hit higher speeds without bogus throttling, it has to be turned off. Generally this is accompanied by an after-market HSF for superior cooling, of course.