make -j lots building and testing gcc

[Bill Northcott <w dot northcott at unsw dot edu dot au>]

On 09/02/2005, at 2:20 PM, Mike Stump wrote:
> > A few early failures:
> > FAIL: gcc.c-torture/compile/20001226-1.c  -O0  (test for excess 
> > errors)
> > Excess errors:
> > cc1: out of memory allocating 457288 bytes after a total of 133158832 
> > bytes
>
> > PS The computer was building another version of the compiler using 
> > 'make -j 20' for some of the time the testsuite was running.  This 
> > was maxing out all 8 cpus and the 8GB of DRAM.
>
> If this anything more than you running your machine into the ground?
>
> Does -j1 work?  If it does, just use it.  In general, you don't want 
> to use more than about -j4 or -j8 I suspect.
>
> If you want to work on gcc to scale it up nicely to large numbers of 
> CPUs, we can point you in the right direction...  let us know.
The computer is an Alphaserver 8400/625 with 8 cpus, 8GB of DRAM and 
about 3TB of attached storage. (coming up to its 8th birthday!)
For the record from 'top -u':
When building gcc with with 'make -j 20 bootstrap'  most cpu time is 
spent in cc1 or cc1plus.  Up to twenty instances of these compilers 
seem to run without problem.  Each instance gets either 33% or 50% of a 
cpu less some miniscule OS overhead.

When running the test suite with 'make -k -j 20 check'  there is never 
more than one instance of cc1 or cc1plus running although there may be 
up to 4 instances of expect.

I think make (v3.80) does what it says on the tin.

Bill Northcott