On Thu, Aug 25, 2011 at 9:16 AM, Dale <rdalek1...@gmail.com> wrote: > Michael Mol wrote: >> >> Sometimes. My understanding is that it comes from packages which are >> badly constructed, and can't reliably handle parallel builds. I'm told >> that these cases are bugs and should be reported. Sometimes, if I >> watch build output fly by, I'll even see something like >> >> make -j10 -j1 (some target name) >> >> where a maintainer decided to put an overriding -j1 after MAKEOPTS. >> >> >> About two years ago, I found that, on my system (quad-core AMD Phenom >> 9650), -j8 resulted in the fastest build time, as measured by building >> ffmpeg.[1] Currently, I'm running -j10, and that's because I've been >> using distcc to pass a couple compile tasks off to other systems. >> (Though with the box I was deferring to scrapped for parts, I'll drop >> this down to -j8 again) >> >> [1] Tested by building in tmpfs. You can find my data here: >> >> http://multimedia.cx/eggs/ffmpeg-and-multiple-build-threads/#comment-150325 >> >> > > This is interesting. I changed mine to j8 too. See how this works. CPUs > are so fast nowadays, they can do a lot pretty fast. I'm trying to imagine > 10 years from now. O_O
I imagine there's a *lot* of money to be made researching AST optimization and simplification transform searches as delegated to GPUs. I don't imagine it'll speed up compiling much, but I do imagine the resulting optimized programs will run far faster. Especially if there's a trend toward declarative programming languages. (I wonder if there's a Prolog implementation that dispatches work to OpenCL. I'm surprised the Mozart/Oz folks haven't dug in that direction yet.) -- :wq