On Thu, 2025-06-26 at 19:21 +0200, wrotycz wrote: > > > There are plenty of scenarios where using more jobs than > > > processor threads results in faster builds: it all depends > > You say that because you have tested it or because you believe it? > I have tested it, But let's bust this ludicrous idea and show us a > test that disproves me. Here is slightly improved test script
I don't see any need to be aggressive in your response. Yes, I have tested it. I have a large C program build using non-local storage that has high bandwidth (can load multiple files in parallel) but also high latency (satisfying any specific read is slow). I can get -j almost 2x the number of CPU threads before builds start slowing down, because (a) the amount of CPU time needed to compile a C program (in relation to other languages like C++ or Rust etc.) is relatively little compared to the size of the source code, and (b) there's a lot of latency of compilers waiting for files to be read over the network. Of course none of this is very relevant to the issue under discussion. > > > However, -j without any limit works the way it does because it is > > > intended to be used in conjunction with -l. > > I don't necessarily use it as I tried and have not seen any > difference, but if you explain what it does I can prepare a test to > actually examine it. The -l (--load-average) option is described in the man page and in the documentation: https://www.gnu.org/software/make/manual/html_node/Options-Summary.html If you run "make -j -l 4" for example it will run enough parallel jobs to keep the CPU load at 4, more or less (basically, keeping 4 CPUs busy). If you use -l without -j, it's useless because you've not enabled any parallelism. That's the purpose of using "-j" with no argument: it means run as many jobs as you can _subject to other constraints_... in this case load level. I'm not saying that it's the most amazingly designed interface, I'm just explaining why -j with no arguments exists and works the way it does. One could imagine an interface like this, which might be easier to use: * -j by itself means the same as -j$(ncpus) or something similar. * -l automatically enables "full parallelism" (what "-j" does today) as a side-effect. * Both -l and -j limits by both number of threads and load average. * If the current "-j" is useful without "-l", add an option like "-j+" or something which enables it. But, that's not how GNU Make currently works. Whether it should be changed is something to be discussed. Changing the command line interface of a 30+ year old program which is invoked probably hundreds of thousands of times a day all over the world, is not to be taken lightly.
