Re: RBTree implementation and thread structure

On Mon, Apr 14, 2008 at 10:56 AM, Axel Dörfler <[email protected]> wrote:
>  See http://people.freebsd.org/~kris/scaling/7.0%20Preview.pdf for
>  example. Of course, that might not be a benchmark we want to excel in
>  :-)


Nice paper. BTW, the ongoing work on FreeBSD 8-CURRENT already shows
loads of improvements upon the ULE scheduler present in 7.0-RELEASE.


On Mon, Apr 14, 2008 at 1:41 PM, Rene Gollent <[email protected]> wrote:
>  True :) It's much more difficult to benchmark interactive
>  responsiveness though, that almost requires a double blind usage test.
>  At least I can't think of any good synthetic way to measure that
>  meaningfully :)


I'm going to state the obvious here, but... Heuristics fail, otherwise
they wouldn't be heuristics. I still have the strongest conviction
that interactivity should be flagged according to the code path being
followed, and not some assumption based on time before blocking/CPU
cycles used. This way the decision can be truly fine-grained and with
zero code overhead on the scheduler itself. IMHO touching parts of the
Input Server and App Server (and perhaps a couple others) to add a
single flag to the semaphores created by them is a very small price to
pay to have perfect interactivity information. Thank God we have the
code for the whole enchilada (and notice how the Unices go to the
craziest lengths to detect and boost X11-like behaviour -- I'm
affirming we're in the very privileged position of *not* needing this
at all).

I know Axel is still unconvinced, but... I already volunteered myself
to work on this after the basic scheduler foundation is laid. Can't
hurt to try :)


Cheers,
A.

-------------------------------------------------------------------------
This SF.net email is sponsored by the 2008 JavaOne(SM) Conference 
Don't miss this year's exciting event. There's still time to save $100. 
Use priority code J8TL2D2. 
http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone