Re: Interesting paper from Perdue

[Vladimir Makarov <vmakarov@redhat.com>]

Steven Bosscher wrote:

>I don't know if anyone has ever seen/read/mentioned this paper
>before, I might have missed it.  Otherwise, interesting reading:
>https://engineering.purdue.edu/ECE/Research/TR/2004pdfs/TR-ECE-04-01.pdf
>
>  
>
The most interesting thing about the article is that they spent a lot of 
machine time (which I have no in my disposal) to investigate individual 
options to get a better SPECInt2000 results.

But I see they used a black box approach because they don't know gcc 
internals at all (they tried -fschedule-insns for p4 which does nothing, 
they also did not use -mtune=pentium4, etc).

Their most complex algorithm (3rd algorithm) to choose better option 
combination is just oversimplified taboo search algorithm (with list of 
taboo moves which never expire).  I think that an algorithm based on 
taboo metaheuristic would achieve better results for the same number of 
tries.  Imho the taboo algorithm is the best fit approach for solution 
of the task (genetic apporach used by Scott Ladd or more random 
semulated annealing approach would work much worse on my opinion).

In any case, the approach is not practical (on my evaluation it needs 
about 15 hours to choose options by the 3rd algorithm for one 
SPECInt2000 test -- three 3 minutes runs, 20 options, 4 iteration as 
they reported).  Alhough it could be used to get a better (peak) 
SPECInt2000 report.

Vlad