Explore solutions for 'portable' performance monitoring with ReFrame

[casparvl]

ReFrame supports performance checking, but this requires hard-coding system name & expected performance in the tests. This breaks test portability. Some ideas would be:

• Use the standard ReFrame way of comparing against expected performance, but place actual performance references in some file external to the test. These would be system specific, but would at least allow performance checks on known systems, e.g. the ones used in EESSI's own CI.
• Compare results to known results, with some very large margin. At least this could capture complete mistakes.
  Implement support in ReFrame for outlier detection, instead of hard performance references. This could check e.g. if the current result is within +/- 2 SD of the past 100 results.
• Implement support in ReFrame for modelling expected performance (e.g. predict expected performance based on core counts, clock speed, processor model...), and then compare to that (again with large margin).
• Do performance checks externally, based on the ReFrame report.

This task should explore different options, then decide on how to proceed (e.g. maybe we go for option (1) in the short term, but work on other long term options?). Since we planned on co-design with ReFrame, those options are also attractive.