Hello Simon, zimoun writes: > Hi Maxim, > > Thanks for the review and the improved patch. > > I am sorry if the commit message and/or changelog I provided was badly > worded, but somehow it was an attempt to explain the odd behaviour – at > least counter-intuitive since I initially felt into when sending the > very first patch allowing parallel tests and you felt too, I guess. :-) No worries. Communicating changes (or anything) is always one of the greatest challenges in programming and elsewhere, it seems :-). The nice thing about it is that it can be improved with perseverance and some feedback. > > On Fri, 26 Nov 2021 at 22:17, Maxim Cournoyer wrote: > >>> --- >>> guix/build/julia-build-system.scm | 8 +++++--- >>> 1 file changed, 5 insertions(+), 3 deletions(-) >>> >>> diff --git a/guix/build/julia-build-system.scm b/guix/build/julia-build-system.scm >>> index f0dc419c17..af478fd4a3 100644 >>> --- a/guix/build/julia-build-system.scm >>> +++ b/guix/build/julia-build-system.scm >>> @@ -112,7 +112,10 @@ (define* (check #:key tests? source inputs outputs julia-package-name >>> (builddir (string-append out "/share/julia/")) >>> (jobs (if parallel-tests? >>> (number->string (parallel-job-count)) >>> - "1"))) >>> + "1")) >>> + (nprocs (if parallel-tests? >>> + (string-append "--procs=" jobs) >>> + ""))) >>> ;; With a patch, SOURCE_DATE_EPOCH is honored >>> (setenv "SOURCE_DATE_EPOCH" "1") >>> (setenv "JULIA_DEPOT_PATH" builddir) >>> @@ -122,8 +125,7 @@ (define* (check #:key tests? source inputs outputs julia-package-name >>> ""))) >>> (setenv "JULIA_CPU_THREADS" jobs) >>> (setenv "HOME" "/tmp") >>> - (invoke "julia" "--depwarn=yes" >>> - (string-append "--procs=" jobs) >>> + (invoke "julia" "--depwarn=yes" nprocs >> >> Here nprocs can be ""; is it really OK to pass an empty string argument >> to julia? > > Yes it is OK. When #:parallel-tests? sets to #f, my patch leads to > the call “julia --depwarn=yes” which is valid. Your modified patch > adds another test but leads to the same call “julia --depwarn=yes”. No, it would invoke julia with the following argv list: "julia" "-depwarn=yes" "" [...]; My point is that invoke is equivalent to doing an execlp system call; and the arguments get passed as a list (including that empty string argument when parallel-tests? is #f). Whether this works or not is up to the application, so I'd suggest not relying on it. Consider for example: --8<---------------cut here---------------start------------->8--- (execlp "python" "python" "" "-c" "print('hello')") /gnu/store/cwqv4z5bvb5x6i0zvqgc1j1dnr6w9vp8-profile/bin/python: can't find '__main__' module in '/home/maxim/src/guix-core-updates-next/gnu/packages/' --8<---------------cut here---------------end--------------->8--- It fails because it interprets the empty string argument as the current directory, apparently. If that works with the above Julia invocation, that's great, but it doesn't make it cleaner in my opinion :-). > + (job-count (if parallel-tests? > + (parallel-job-count) > + 1)) > + ;; The --proc argument of Julia *adds* extra processors rather than > + ;; specify the exact count to use, so zero must be specified to > + ;; disable parallel processing... > > [..] > > + (apply invoke "julia" > + `("--depwarn=yes" > + ,@(if parallel-tests? > + ;; XXX: ... but '--procs' doesn't accept 0 as a valid > + ;; value, so just omit the argument entirely. > + (list (string-append "--procs=" > + (number->string additional-procs))) > + '()) > > > So because of 2 tests. I think your modified patch is more > “complicated”. :-) It is slightly more complex indeed; but I think it provides the reader with useful knowledge of julia's quirks and is more correct. > About this, > > + (additional-procs (max 0 (1- job-count)))) > > I considered that it was not a big deal; initially, I did something > similar in ’let’ but remove it because it changes nothing from my > experiments. In fact, because ’--procs’ overrides JULIA_CPU_THREADS and > run Julia with n or n+1 is more or less the same for the Julia land, > IMHO. Well, it is not clear what is the load for the main worker. And > JULIA_CPU_THREADS=1 is required for running using only one worker. > Anyway, this changes nothing, practically speaking. :-) Indeed, it is > better and more consistent. Yeah, I don't like that the behavior of --procs is to *add* workers, rather than set its exact count; which make thing awkward. Even upstream get tricked by that by erroneously *adding* Sys.CPU_THREADS workers because of this in test/runtest.jl [0] [0] https://github.com/JuliaLang/julia/pull/43217#pullrequestreview-817102530 Thanks, Maxim