SingularSpectrumAnalysis and CPU load

Specific Domains Machine Learning
1

SSA does a good job of extracting trend and seasonality from a time series. However widening the window on a large dataset puts quite a load on my CPU (GPU is too old to contribute to calculation in Julia) heats up my CPU quite noticeably. A window of 20 on a data series of 20000 goes quite nicely and finishes without stressing the CPU, but 200 and particularly 2000 window puts a strain on my old i5-4460. So I use watch with sensors and kill the process if it gets too hot.

Flux has a useful callback feature in the train! function which permits building in a check for a hot CPU which is very useful to ensure that you keep the state intact. Ensuring that data is in Float32 format also helps.

I recently renewed the paste on the processor and reseated the fan so I know that is good, and I don’t want to limit the assigned cores; what other measures possible to keep from melting down?

2

Your cpu should really handle that by itself by throttling when it gets too hot. Have you overclocked the cpu or something?

3

No, I have not deliberately overclocked the CPU, but I am possibly being over-careful. “sensors” reports that 80C is hot and critical is 100C+, so I start to worry when the CPU goes above 70. Indeed the CPU will shut itself down at the critical level but I don’t really want a disorderly shutdown, I just need to get to the end of the epoch or loop by allowing the CPU to cool for a while and then continue. And then there is the matter of speeding the eventual failure of the CPU by running it hot.

4

I don’t think you need to worry about the temperature then, the cpu will reduce its clock frequency to keep the temperature below critical, I’ve never experienced a fatal error due to overheating of a modern cpu.