Re: Parallel Compilation with gnustep-make

[David Chisnall]

On 23 Feb 2009, at 12:48, Fred Kiefer wrote:

> I tried your first version, where you had to ask for the parallel  
> build
> more verbosely, and even on my single processor machine it appeared to
> be about 10% faster. This could have been an error in my measurement,
> though.

The general rule of thumb for parallel make is either number of CPUs +  
1, or number of CPUs * 1.5, depending on who you ask.  In a sequential  
make, you have a lot of time when the CPU is idle, waiting for I/O,  
typically at the start and end of each compilation step.

With a parallel make, you want enough floating processes that, when  
one is waiting for I/O, another can be given CPU time.  For GNUstep  
programs, which are often composed of short, easy-to-compile, sources,  
you may find -j3 or even -j4 is faster on a single-processor system.   
Doing a sequential make of EtoileFoundation on my (slow) machine gives  
these numbers:

real    0m13.089s
user    0m5.777s
sys     0m5.862s

This means that the CPU spent 1.45 seconds, or around 11% of the time,  
either doing unrelated things (unlikely - nothing else was running) or  
waiting for I/O during this build.  The maximum speed-up I would  
expect from parallel compilation here is around 10%.  With more CPUs,  
faster CPUs, or slower disks, the speedup will be more.

David