Ignore noisy points
Byte-Lab opened this issue · comments
I think we should figure out some threshold for ignoring noisy data points. This may not be as big of an issue when profiling long running processes with long experiments on isolated machines, but in general this is a flaw in the profiler. See this graph as an example.
Quick hacks can be taking an average of the points for a given line and speedup and excluding any point that lies outside some epsilon (say 40%).
Fixed in @4198d1a2e211144c8b2b36be418635863cf91d49