[Gcl-devel] prefetching GBC sweep

[Camm Maguire]

Greetings!  After all this GBC work, I wanted to note here an idea
which should be easy to do, but might best be put off.  And that is to
add prefetch instructions to the sweep portions of the GBC to speed
things up.  

Adding the prefetch itself is almost trivial, but to do it right, I
think there should be some autotune process whereby the system
measures the time taken per page at a given prefetch offset, and then
adjusts this offset according to some minimization algorithm, maybe
even a simple parabolic fit with an exponential backoff.  Likewise,
these instructions vary by subarchitecture, i.e x86 binaries would use
a different call on an AMD vs. an Intel, and no call at all on a pII,
etc.  The best way to handle this is IMHO to have a little shared
library with the calls, different versions of which would be built
(i.e. at least three on x86) and installed in such a way that the
system ld.so loader would choose the best version at runtime according
to the running CPU.  Atlas on Debian does this already with some
success. 

Anyway, I don't really know much about how this might work outside the
Debian world, or even what type of gains could be had.  If its only
marginal it won't be worth the effort, but I've read a bit about big
wins here.  Does anybody have relevant information on this?

Take care,

-- 
Camm Maguire                                            address@hidden
==========================================================================
"The earth is but one country, and mankind its citizens."  --  Baha'u'llah