hardware counter-based performance modelling

[Nichols Romero]

Hi,

This is not a question about LIKWID per se, but since I am using LIKWID I thought I would leverage the community of smart people on this list :)

I am interested in building simple analytical performance models for some proxy apps (possibly also applications, depending on how things go) so that I can input a memory bandwidth, network injection, flop rate, etc. and get some reasonable extrapolation for a new architecture. (I am trying to avoid simulators for now). 

LIKWID comes into the picture by helping to characterize the proxy app and the workload on current architectures. 

After doing a literature search, I was able to find some approaches but they were very specific to a single compute kernel (e.g. finite different stencils or FFTs). I am looking for something that could be applied to a proxy app (which would include many compute kernels and computational motifs, not just a single one).

I stumbled upon this paper:

https://link.springer.com/article/10.1007/s00521-018-3496-z

which is kind of what I am looking for. They didn't use LIKWID as far as I can tell, but do not hold that against the authors :)

Is anyone aware of other papers that might be relevant to this kind of hardware counter-based/-informed performance modelling?

Thanks,