Re: [Valgrind-users] Expected slowdown

[Christoph B. <bar...@or...>]

Am Dienstag, 24. Oktober 2006 22:27 schrieb Julian Seward:
> > > hash = (hash << 61) | (hash >> 3);
> > >
> > > leads to a better distribution of the values and better runtime for the
> > > testcase. However I do not know whether rotation by 3 really leads to a
> > > good hash function for 64 bit.
> >
> > Yeh, rotating by 3 is stupid and doesn't mix the bits well for short
> > stack traces.  I just tried this, which mixes much better:
>
> Below is my final patch.  It works well (it seems) for both 32-bit
> and 64-bit versions of your test program and contains a second
> heuristic hack to guard against bad-case behaviour.  Can you let me
> know if it solves the problem convincingly?  If so I will commit it.

The main application runs faster now. One run of the algorithm:

Old valgrind: 144765 seconds
New valgrind:  17232 seconds
Without valgrind: 251 seconds

Although this is nearly ten times faster than the old version it is still 
quite slow for me with a slowdown factor of 68x. But I guess that memcheck 
cannot run faster.

Thanks
Christoph Bartoschek