> I'd like to rework the current sysfs cpu/node pieces to provide
> two separate topologies (one for CPU and one for memory).
There are two other sets of patches that are floating around that
will interact with this. The first (from Matt) makes the topo
functions cache results, instead of recalculating each time.
I'm thinking that we may want the generic front end to do that and
read from arrays, whilst only the *_init backend gets moved into
the subarch stuff ? (looking back at the patch right now, I can't
quite see what it's doing in 30s, but that's what it was meant to do).
The second is the start of breaking things out into the numaq
subarch - the attatched bits seem to work & are written with the
intent of being mergeable. It's just pieces torn off the big patch
by James Cleverdon and John Stultz (which isn't really mergeable
as is) and tweaked around by me. I'm going to test them a little
more before submission, but they're pretty much there, I think.
They're based on top of John's subarch reshuffle (the stuff you had
under generic is really NUMA-Q)
> Ultimately, the scheduler could be tuned to use the topologies to
> make scheduling decisions.
We kind of have that already. There are NUMA sceduler patches that
do this kind of thing in 2.5.47-mjb3 ... earlier versions of Erich's
scheduler had a pooling abstraction - this got ripped out for
simplicity in the hope of getting something merged before the 2.5
freeze, but it'd be nice to put them back if that's not going to
happen (still vaguely hoping).
> When that happens, we can probably fold
> the current Pentium Hyperthreading stuff into a simple topology map
> as well.
Not sure about that ... the HT stuff I believe created one queue
per pair of CPUs, which isn't going to work well for multiple real
CPUs per nodes, though is kind of a nice trick for a shared cache
SMT like HT .... things like the PPC64 chip multi-chip-on-1-die
may feel differently about that.
> I believe Martin Bligh and Bill Irwin are working (or at least
> thinking) somewhat along these lines, so I thought I'd gather
> feedback before jumping into a wholesale rewrite.
cc'ed John, Erich, Matt ...
M.
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Sat Nov 30 2002 - 22:00:23 EST