Julia 1.1 is what we’re still testing on since 1.2 came out last week. I don’t know if there’s a performance regression there.
Anything on the saving?
Julia will not necessarily deallocate the cache from the first or second run before starting the third. I know from running your script locally that the third time more than doubles if you the GC of the third run is counted as in the third run, so you might want to GC.gc() so the runs don’t have that cross talk.
Also, can you share the script you are running?