Attached is my simple NUMA scheduler patch. It has been split into
two separate patches, the first part is a modification to
find_busiest_queue to favor runqueues for processors on the same
node. Care was taken to ensure that the change to find_busiest_queue
would not change the selection criteria used for non-NUMA machines.
The second patch adds logic to exec to place the exec'd task on
the least loaded runqueue, thus avoiding the need to move the task
to a different (potentially off-node) runqueue in the immediate
future. This patch needs the first patch to apply, but could be
easily changed to apply to a kernel without the first patch.
There are no functional changes between these two patches and the
patch provided for 2.5.43. As stated then:
On NUMA systems these two changes have shown performance gains
in the 5 - 10% range depending on tests. Some micro-benchmarks
provided by Erich Focht which stress the memory subsystem show
a doubling in performance.
The simple NUMA scheduler patch has been included in the
linux-2.5.44dcl1 kernel so has seen additional testing with no
issues reported. I would welcome additional testers or inclusion
in other 2.5 kernel trees.
--
Michael Hohnbaum 503-578-5486
[email protected] T/L 775-5486
On Wed, 2002-10-23 at 10:53, Michael Hohnbaum wrote:
> Attached is my simple NUMA scheduler patch. It has been split into
> two separate patches,
Here is the second piece of the NUMA scheduler patch. This patch adds
logic to exec to place the exec'd task on the least loaded runqueue.
This only effects kernels built with CONFIG_NUMA.
Please consider for inclusion in 2.5 kernel trees.
--
Michael Hohnbaum 503-578-5486
[email protected] T/L 775-5486