On Fri, Feb 21, 2020 at 6:13 PM <[email protected]> wrote: > > 1.8: > > real 3m41.714s > > user 6m52.414s > > sys 0m36.662s > > > > 2.2: > > real 6m8.344s > > user 12m51.799s > > sys 0m34.875s > > > > I think this warrants some deeper look, though: I have the impression > that the > > lilypond-book regression tests aren't parallelized. They also suffer > from the > > higher start up costs (there are some 60 of them, so that's 90 seconds > right > > there). > > I recently found some "could not contact jobserver" in the logs from > make and googling it suggested that this is a consequence of using > $(MAKE) rather than $$(MAKE) inside of macros (causing premature or too > much expansion?). I have not investigated yet and don't know what > counts as "macro" here. Or whether this is all nonsense. Either way, > it should affect Guile-2 just as much as Guile-1.8, shouldn't it?
The startup costs of parsing .scm files of 1.5 secs gets added to the process 60 times. If you remove that, >>> (368 - 90) /221. 1.257918552036199 we're back to a 25% slowdown. > > https://codereview.appspot.com/569390043/ -- Han-Wen Nienhuys - [email protected] - http://www.xs4all.nl/~hanwen
