Please [post the patch] ... It probably needs to be thread-safe, and I vaguely recall that lex/yacc output isn't.
In a couple of weeks it will be properly tested and, as you pointed
out, I still need to switch to bison/lex's reentrant parser. I'll post
it on the mailing list once those things are ironed out.
I've followed the coding standards but there are for a few things that
need to be fixed. The biggest one is that I was unsure where to put the
new code so I put it in a new library - libexpr, but then anything
that links against libsched (containing scaled_mixed_load(), and its
call to get_load_value(), which I now nest in
if (mconf_get_use_new_load_formula_evaluator()
...
else
...'
) or libsgeobj (containing validate_load_formula() and its call to
centry_list_locate(), which I now nest ...) now has to also link against
libexpr. So the patch contains a *lot* of changed Makefiles. If you have
any advice on where to place the code then it would be appreciated! Also,
I could so with shortening the name of the config directive that enables
this feature :-) For now, and in case it helps solicit feedback, I've
posted it at http://pastebin.com/39yaMQUE.
It would be good to explain the motivation and why the current parser isn't good enough (I don't remember how it works and whether it's used elsewhere),
Apparently the reason is that some of our customers are using a tool
that has a strong preference for using whole physical cores rather than
(hyperthreaded) virtual cores. E.g. on a system with 8 CPUs each with
4 cores each with 2 virtual cores, then the load formula
(32-slots)^2
would encourage the scheduler to place the next place-able job on
a node which was the closest to running a job on each physical core.
Okay, it doesn't express that a machine running 31 jobs is preferable
to one running 33 jobs, but that could be accomplished with some
quartic polynomial.
Alexis
<<attachment: alexis_huxley.vcf>>
_______________________________________________ SGE-discuss mailing list [email protected] https://arc.liv.ac.uk/mailman/listinfo/sge-discuss
