[Linux-ia64] node affine NUMA scheduler

From: Erich Focht <efocht_at_ess.nec.de>
Date: 2002-07-22 23:59:41
There's a new version of the node affine NUMA scheduler extension based
on the O(1) scheduler at

The patch is for 2.4.18 kernels and it has been tested on IA64 systems.
It requires the O(1) scheduler patch with the corrected complex macros
which I posted to the LSE and linux-ia64 mailing lists last week. For
IA64 you should use:
which should be applied to   2.4.18  +  ia64-020622 patch.
For IA32 (NUMA-Q) try instead:

What is it good for?

 - Extends the scheduler to NUMA.
 - Each task gets a homenode assigned at start (initial load balancing).
 - A memory affinity patch (like discontigmem, or similar) should take
   care that the memory of the task is allocated mainly from its
 - The scheduler attracts the tasks to their homenodes while trying to
   keep the nodes equally balanced.
 - Target: keep processes and their memory on the same node to reduce
   memory access latencies without having to fiddle with the
   cpus_allowed masks (hard affinities).
 - Within one node, behaves like the normal O(1) scheduler.

For an overview over the features have a look at:

There are several changes compared to the previous version, the most
important ones are:
 - Extension to multilevel NUMA hierarchy by implementing delays when
   stealing tasks from remote nodes.
 - Better selection of task to be stolen from busiest runqueue. Take
   into account cache coolness, node and supernode of task and runqueue.

Comments and feedback are very wellcome.

Received on Mon Jul 22 07:00:18 2002

This archive was generated by hypermail 2.1.8 : 2005-08-02 09:20:09 EST