Gentoo thread post-mortem
Greetings! I just saw the Gentoo thread reported on DWN. Please
excuse the lateness of this suggestion -- I hope it will be worth your
attention.
In my opinion, we should make sure to get the *very large* performance
gains (e.g. ~ >=50%) available through subarch package tuning, and
leave the 5-20% gains for later if we ever have time. In my
experience, such gains are available in a very few well-defined areas
of computation -- floating point linear algebra, perhaps random
number generation and hashing, multiple precision arithmetic, etc. I
think that we should
1) identify no more than 20 such areas
2) isolate the performance critical code in shared libraries, and
3) rely on the dynamic loader and its configuration to load the most
optimal version at runtime.
blas/lapack/atlas already does this, to great performance benefit,
which is then immediately available to all programs linking against
blas/lapack without further re-compilation. there are currently quite
a few of these programs in Debian.
I put this together into a policy proposal last year, but it seems to
have foundered:
http://bugs.debian.org/cgi-bin/bugreport.cgi?bug=120418
Comments/action on this proposal would be most appreciated!
Take care,
--
Camm Maguire camm@enhanced.com
==========================================================================
"The earth is but one country, and mankind its citizens." -- Baha'u'llah
Reply to: