Re: Questions, a patch and a script

David Landgren wrote:

>> F<configsmoke.pl> offers you the {makeopt} variable for the "make" and 
>> "make test-prep" steps. It also offers the {testmake} variable, which 
>> defaults to $Config{make}, but you could try to set that to "make -j" 
>> (but I haven't tested that).
>> This deliberate split between makeoptions for different steps in the 
>> process was done because Alan ran into problems running the equivalent 
>> of "make -j _test", but remember parallel "make test" it not 
>> guaranteed to work!
> 
> Why is that? My latest servers are Xeons with the so-called HTT stuff. I 
> noticed that smoking 5.8 or 5.9 took 23 minutes per configuration on a 
> standard kernel on FreeBSD. When I recompiled the kernel with HTT 
> support, that time went down to 17 minutes.
> 
> But now my system load maxes out at about 0.48. If parallel make tests 
> were guaranteed to work, the process would be able to soak up more 
> cycles going to waste. Are there archives on the subject (of prblems 
> with //l makes), either here or on p5p?

There are race conditions between some of the tests in the perl test suite, 
i.e. one test initialises files used by 'later' tests - except that with 
parallel make 'later' may not actually be later.  I forget which exact ones 
were the problem, but feel free to fix! ;-)

-- 
Alan Burlison
--