Re: [boost] Is a 3% timing difference reliably repeatable in real code?

From: Darren Cook (darren_at_[hidden])
Date: 2003-05-08 18:23:34

>> A 2-3% timing difference probably isn't reliably repeatable in real code.

My timing tests were interleaved as I swapped in various algorithms and 
recompiled and the time differences were consistent. I gave ranges not 
averages to give a feel for the "experimental error".

> I'm not sure, though, if this negates your point, Beman. Something that 
> gives a 2-3% speedup for one Boost user might not be worth any level of 
> obfuscation unless we can prove it provides a similar speedup for other 
> Boost uses.

Just to avoid confusion for people joining this thread halfway, the 2-3% 
speedup was in my code; I wasn't proposing obfuscating any boost library.

My point was that I got this speed-up by overriding operator new/delete with 
one-liners in just a single class (of course I chose the class I new from 
profiling was making heaviest use of new/delete):

void* operator new(std::size_t sz){
   return boost::detail::quick_allocator<T>::alloc(sz);
   }

void operator delete(void *p,std::size_t sz){
   boost::detail::quick_allocator<T>::dealloc(p,sz);
   }

where T is the name of my class. That is so easy it is practically free, 
like using -O3 instead of -O2.

 > In our project, a 3% speed increase is considered a cause for celebration.
 > ....
> - sped up JavaScript iBench by 70% by using a better sort algorithm and 
> reducing the number of UString allocations

How did you celebrate that one then? :-)

Darren