performance tuning

[Russ Allbery]

Katsuhiro Kondou <kondou at nec.co.jp> writes:

> Innd consumes CPU time in processing ARTclean(), which Barry already
> noted before.  I probably find where it is.  Data copying(*out++ =
> *in++;).  This is done for all incoming articles.  And I found there are
> two other places to data copy.  One is end part of NCproc(), and the
> other is ARTstore().

> I think I can fix this, but I'd change SMstore() interface to use
> writev() style, and some of processing in ARTclean() need to be done in
> NCproc().

> Any comments?

I don't have any problems with this; I'd been thinking for a while that
this might be a better approach.  I think we could recover a chunk of
performance this way.  We *do* need to have the body of the article as a
contiguous chunk of memory so that the filtering interface will work
right, though.

-- 
Russ Allbery (rra at stanford.edu)             <http://www.eyrie.org/~eagle/>