Danek and I both noticed that, with 200,000+ actions on a system, action
creation (from string --> action object) is actually showing up in our
performance profiling. It's not a massive problem, but it's there.
Here's my attempt at a performance rewrite. On my AMD64 system I'm
seeing the time it takes to create 100,000 actions drop from about
7.7secs to about 4.1secs. So we're up to about 25,000 actions/sec.
The key was to accumulate the action into dictionary all at one, instead
of using intermediate lists.
I also fixed a number of irritating bugs in the parsing, and wrote
a simple benchmarker, and a set of test cases.
Bart and I both agree that if we need more performance, the next
step will be to write action_string -> attribute_dictionary in C--
I would guess that we can probably get at least another doubling.
The pareser is still not perfect-- but it's better than it was. I
think it'll be hard to keep things fast and make them more correct
than this, without going to C.
Your comments welcome.
http://cr.opensolaris.org/~dp/ips-actions/
Thanks,
-dp
--
Daniel Price - Solaris Kernel Engineering - [EMAIL PROTECTED] - blogs.sun.com/dp
_______________________________________________
pkg-discuss mailing list
[email protected]
http://mail.opensolaris.org/mailman/listinfo/pkg-discuss
