[lkml]   [2002]   [May]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectNode Affine NUMA scheduler, updated

an updated patch for the node affine NUMA scheduler extension based on the
O(1) scheduler can be found at

Detailed information on the implementation is at .

What's new:
The topology information has been updated and supports the following ccNUMA
- IBM NUMA-Q - i386 (thanks to Matt Dobson),
- SGI SN1/2 - ia64 (thanks to Jesse Barnes),
- NEC AzusA - ia64
No other i386 platforms have been tested, yet.

The topology info now uses the notions of logical and physical node, also the
variables are protected by a rw_lock. This was a must for the integration
with the cpu-hotplug patch.

There are two configuration variables which control the way how the scheduler
- CONFIG_NUMA_SCHED : switch on pooling scheduler (otherwise it behaves like
the O(1) scheduler, though it looks different).
- CONFIG_NODE_AFFINE_SCHED: tasks remember their homenode and are attracted
back to it.
For platforms with a big node-level cache it might be better to only configure
better if the penalty for trashing the node-level cache is bigger than the
benefit of running on the right node (where the memory is allocated).

I added a variable node_policy to the task structure which is inheritable and
decides on the initial load balancing. There is a prctl interface to change
this from userland, a utility called nodpol is available on the web page. The
possible values for node_policy are:
0 (default) : select homenode in do_exec(),
1 : select homenode in do_fork() only if CLONE_VM is unset,
2 : select homenode in do_fork() (allways).
It's mainly meant for experiments and benchmarks. Some benchmarks (e.g. AIM7,
which simulates large loads) only fork but don't exec, thus the default
homenode selection mechanism doesn't apply and the load balance is bad right
from the start. In real life one should just check whether multithreaded jobs
need to be distributed across multiple nodes or better take their memory from
one node and change the node_policy accordingly before starting them. The
default behavior should be fine otherwise.

On the web page I included some results showing performance increase with the
node affine scheduler and its functionality. Basically it works fine for
medium and high loads but has some trouble with low loads. This is due to the
fact that a task running on a remote node alone on its CPU cannot be stolen
by CPUs on the homenode. load_balance() is called in such places that the
only mechanism for moving a currently running task (migration_thread) cannot
be used. Any ideas (besides a signal) are welcome. The initial load balancing
is improveable, too, a better measure for load will help.

Thanks in advance for your feedback, I'm especially curious about results for
the affinity_test on other platforms than NEC AzusA.

Best regards,
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to
More majordomo info at
Please read the FAQ at

 \ /
  Last update: 2005-03-22 13:25    [W:0.032 / U:3.496 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site