On Sat, 30 Sep 2000, Andrea Arcangeli wrote:
> On Sat, Sep 30, 2000 at 02:58:35PM -0300, Rik van Riel wrote:
> > Both the current code and classzone will need some adjustments
> > to work fine (as in: reasonably efficient) with NUMA.
>
> The exact thing I was talking about in my previous email is that
> in classzone _all_ the cache lru are been intentionally designed
> to be per node lists because NUMA memory balancing heuristics
> needs that per-node info to be able to shrink from the "near"
> node.
Yeah, but this is a no-brainer change which can also be
applied to the normal VM in 30 minutes.
The real reason classzone could be interesting is that
(combined with per-node memory_pressure info) it could
be used to better balance memory pressure across nodes.
Say that the local node got allocated a slightly too big
job and the node next door has some spare capacity. In
that case classzone could be used to make sure we'll:
1) use some memory from the other node instead of swapping
2) don't start swapping too much on the local node because
it wouldn't be good
Using information like memory_pressure difference and node
distance, we can use classzone to use the NUMA beast a bit
more efficiently.
The only question here is, can we abstract this away in
such a way that:
1) the code is clean and efficient
2) we don't do any of this complex stuff on "normal" machines
Oh, and then there's of course the question of the pagecache_lock,
but that's quite a separate issue from the memory allocation and
balancing ;)
regards,
Rik
--
"What you're running that piece of shit Gnome?!?!"
-- Miguel de Icaza, UKUUG 2000
http://www.conectiva.com/ http://www.surriel.com/
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Sat Sep 30 2000 - 21:00:27 EST