Hi, Peter Simons wrote:
> Using -j x could lead to different builds (doesn't in most cases though). Um, what exactly do you mean? Are you saying that nix doesn't utilize multi-processor machines because there is some inherent design problem that would lead to broken packages if multiple CPUs are being used during the build?
Indeed, but the design problem is with make and makefiles.
> The only way to do this is adding that option to the builder.sh file or > such. (default buildPhase) Does anyone agree with me when I say that this situation is unsatisfactory?
Yes, it's certainly unsatisfactory, but there is no obvious solution. However, the situation isn't all that bad: Nix has its own "-j" option to perform multiple builds in parallel. So if you're (say) rebuilding all of Nixpkgs from scratch, you do get a huge amount of parallelism. And contrary to Make's "-j" this is guaranteed to be safe since builds don't interfere with each other.
Of course, this is parallelism between builds, not *within* builds, so if you're rebuilding OpenOffice it doesn't help you :-(
> It would be interesting to have an experimental maybe unpure use numcores > environment option which would be 1 in all production cases (buildfarm) > but might be set to another value on your local developement machine. What exactly makes a parallel build "unpure"? Are you referring to broken makefiles that contain race conditions and thus fail when make executes build jobs concurrently?
Exactly. There is no way to guarantee that a parallel Make produces the same result as a sequential Make. But I have no idea how many packages have broken makefiles (my guess is a lot).
OTOH, I'm mostly worried about race conditions that don't cause the build to fail but produce a different result, and that's probably very rare. So turning on "-j" globally might be acceptable, but some empirical results would be needed :-)
(Undoubtedly Gentoo people regularly build packages with specific -j flags, so they probably know how much breakage it causes...)
-- Eelco Dolstra | http://www.st.ewi.tudelft.nl/~dolstra/ _______________________________________________ nix-dev mailing list [email protected] https://mail.cs.uu.nl/mailman/listinfo/nix-dev
