Pádraig Brady wrote:
> On 11/12/10 09:05, Paul Eggert wrote:
>> It seems to me that this code in sort.c:
>>
>> unsigned long int np2 = num_processors (NPROC_CURRENT_OVERRIDABLE);
>> if (!nthreads || nthreads > np2)
>> nthreads = np2;
>>
>> is now obsolete. It was written assuming spin locks, but now that
>> we use mutexes, shouldn't we respect an explicit --parallel=N
>> flag? Something like the following, say? This would let the user
>> override the environment in the command line, which is normally what
>> people would expect.
>
> Yes I think you're right.
> Related to this is the default number chosen,
> which might be best to restrict to 8 or so
> as there are diminishing returns after that.
8 sounds like a reasonable threshold.
I ran tests on a 6-core i7-970 (which shows as 12-core):
gensort -a 10000000 in-10M
env time --format=%e sort in-10M > /dev/null
N-parallel elapsed_seconds (avg of 10 trials)
12 6.91
8 7.13
gensort -a 20000000 in-20M
env time --format=%e sort in-20M > /dev/null
N-parallel elapsed_seconds (avg of 10 trials)
12 13.94
8 14.34
It is indeed a case of diminishing returns, at least in these cases.
> Of course we'd need to benchmark again with
> all the recent changes to find an appropriate default.
> The gcc compile farm has a niagra 32 core (gcc12) and
> and a magny-cours 24 core (gcc10) available.
>
> Also I notice Chen has inadvertently been omitted from THANKS,
I have not been adding commit "author" names to THANKS
partly because eventually I want to generate that file from the
combination of the git commit log and a VC'd list of other
name/email pairs.
> and that his UCLA email address should
> probably be added to .mailmap
Good eye.
