[ltp] Stress testing for undervolting

[Laurent Gilson]

Hello,

> I know that there are two camps of people, those that will tolerate a bad
> calculation, and those that won't.  I'm definitely the later.

Do not undervolt.
Do not use laptops, use servers.
Use ECC RAM.

Every PC makes bad calculations from time to time. For SDRAM the normal  
error-rate is about 1 bit per year. If you want 100% correct operation at  
all times: go server.

> The longest test I've run to
> date has only been about a half an hour, but it's been long enough to  
> prove that 16mV from the cliff isn't enough of a margin away from the  
> cliff.

If you want to find out quickly if you have reached the "safe" side:
- plug/unplug AC and batt at random
- run hdparm random-seek tests on HD and DVD (with pauses)
- and use the Wifi to transfer large amount of data (also with pauses)

The goal is to cause turbulence in the voltage. Quickly going from minimal  
to maximal load and back is more likely to trigger an error then waiting  
for 24 hours...

Also test changing from/to the frequency you are about to adjust.

> Documentation on-line recommends 24 hours of testing for a proposed
> undervoltage setting.

Per frequency. I think 1 hour rough testing is more likely to find an  
error.

> Has anyone
> seen a first failure at, say, the 22nd hour of stress testing?

Yes, but not with a undervolted CPU. It was a bad RAM that triggered  
memtest86 after running 2 days.

> But how much confidence do you guys feel 24 hours of testing
> is really worth?

I have tested 2 hours in total. And added 1 voltage step afterwards. After  
4 months: No hiccups.

cu