From: already5chosen@yahoo.com   
      
   On Thu, 30 Oct 2025 16:04:51 +0100   
   David Brown  wrote:   
      
   > On 30/10/2025 13:07, bart wrote:   
   > >   
   > >   
   > > OK, "make -j" gave a real time of 30s, about three times faster.   
   > > (Not quite sure how that works, given that my machine has only two   
   > > cores.)   
   >   
   > You presumably understand how multi-tasking works when there are more   
   > processes than there are cores to run them.  Sometimes you have more   
   > processes ready to run, in which case some have to wait.  But   
   > sometimes processes are already waiting for something else (typically   
   > disk I/O here, but it could be networking or other things).  So while   
   > one compile task is waiting for the disk, another one can be running.   
   >  It's not common for the speedup from "make -j" or "make -j N" for   
   > some number N to be greater than the number of cores, but it can   
   > happen for small numbers of cores and slow disk.   
   >   
      
   It *can* give much higher speedup than the number of cores.   
   Measurements taken at relatively small MCU project: 33 modules,   
   size:   
      text    data     bss     dec     hex filename   
     26953     156   28028   55137    d761   
      
   Compiled on my corporate desktop.   
   Good hardware (Intel i7-17700, 8 P cores, 12 E cores, 28 logical CPUs,   
   competent SSD : Samsung PM9F1).   
   Bad software environment - very aggressive antivirus + 2 other   
   "management" crapware agents.   
      
   msys2, arm-none-eabi-gcc 13.3.0   
      
   2nd column: execution time with all cores enabled.   
   3rd column: execution time with compilation locked to single   
   logical CPU (P-core).   
   4th column: execution time with compilation locked to single   
   logical CPU (E-core).   
      
   flags tm-all     tm-one-P   tm-one-E   
   none  0m20.689s  0m21.162s  0m44.608s   
   -j 2  0m9.464s   0m11.199s  0m34.154s   
   -j 3  0m6.855s   0m8.695s   
   -j 4  0m4.970s   0m7.992s   0m21.895s   
   -j 5  0m4.429s   0m7.632s   
   -j 6  0m4.016s   0m7.340s   
   -j 7  0m3.766s   0m7.296s   
   -j 8  0m3.564s   0m7.248s   
   -j 9  0m3.439s   0m7.245s   0m20.323s   
   -j 10 0m3.562s   0m7.324s   
   -j 28 0m3.741s   0m7.295s   
   -j 33 0m3.623s   0m7.128s   0m18.098s   
   -j    0m3.843s   0m7.187s   0m19.365s   
      
   So, on P-core I see almost 3x speed up from simultaneity even with no   
   actual parallelism.   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)