Torture test not torture enough?

[cmokruhl]

I've seen some weird behavior that makes me think the torture test isn't strenuous enough to the CPU... After running the torture test error-free for many hours, I went back to testing my exponent, which uses 896K FFTs. I started getting SUMOUT errors and lockups.

To track down what was happening, I started monitoring my CPU temperature. During the torture test, it stabalized and never went higher than 60C. Once I went to normal testing with an 896K exponent, the temperature quickly rose to 61C and stabalized. Clearly, normal testing is stressing the CPU more than the torture test.

The first thing that springs to mind is the way the torture test uses memory. I have 512MB installed and the torture test uses 384MB, whereas normal testing uses ~12MB. Perhaps all the extended memory usage stresses a different portion of the CPU? Or gives the FPU a break?

The torture test seems to run one set of iterations using max memory, followed by one set using the "normal" amount of memory. It's hard to tell, but I think the iterations using max memory are allowing the CPU to cool down. Is there a way to force the torture testing to only use the "normal" amount of memory?

[Deamiter]

yes. When you run the torture test, it automatically asks you how much memory you'd like to use. Default is around 90% I think, but I believe that you can use whatever you want by changing it in the CPU menu.

[NickGlover]

Originally Posted by cmokruhl

I've seen some weird behavior that makes me think the torture test isn't strenuous enough to the CPU... After running the torture test error-free for many hours, I went back to testing my exponent, which uses 896K FFTs. I started getting SUMOUT errors and lockups.

To track down what was happening, I started monitoring my CPU temperature. During the torture test, it stabalized and never went higher than 60C. Once I went to normal testing with an 896K exponent, the temperature quickly rose to 61C and stabalized. Clearly, normal testing is stressing the CPU more than the torture test.

The first thing that springs to mind is the way the torture test uses memory. I have 512MB installed and the torture test uses 384MB, whereas normal testing uses ~12MB. Perhaps all the extended memory usage stresses a different portion of the CPU? Or gives the FPU a break?

The torture test seems to run one set of iterations using max memory, followed by one set using the "normal" amount of memory. It's hard to tell, but I think the iterations using max memory are allowing the CPU to cool down. Is there a way to force the torture testing to only use the "normal" amount of memory?

There isn't a way unless you change the amount of memory in Options->CPU. However, I think there should be a way. Maybe it could be just a single checkbox or even an "undocumented" prime.ini setting that lets you toggle whether it uses just the memory needed for the LL iteration or whether it uses the maximum amount of memory for P-1. Given cmorkruhl's issue, I also think the default setting should be to just use the memory needed for the LL iteration.

[cmokruhl]

I got this reply from an e-mail I sent to George:

In Options/CPU set your available memory to 8MB prior to running the torture test.

Hope that helps,
George

That worked for me. I can't say for certain whether or not the default should be changed to use the minimum amount of memory. Doing so stressed my CPU more, but perhaps using additional memory stresses other CPUs more. Perhaps it could be an option that could be turned on/off instead of messing with the CPU memory options.