Hi,
here come two patches which together implement the pooling NUMA scheduler
with initial load balancing stripped of the node affine scheduler.
First patch does:
- implement CPU pools structure
- changes find_busiest_queue to first scan the own node, then (if no
imbalance found) find the most loaded node and steal a task from it
if the imbalance exceeds 25%.
- The steal is delayed 100ms if we steal from a remote node and the own
node is averagely (within 25%) loaded.
- The steal is delayed 1ms if the own node is unloaded (this gives
CPUs from the own node an advantage).
I dropped the multilevel structure as it is only complicating things at
this stage and can be introduced later. The patch is prepared for it.
The second patch implements initial balancing by selecting the most
unloaded node and its most unloaded CPU for an execed task. It does
a round-robin selection of the next node if the loads are equal.
The node affine patch will be on top of this.
The pool setup is pretty messy, it needs some sort of RCU to become
nicer. Also undefining CONFIG_NUMA_SCHED could lead to errors, I didn't
clean this up. This is mainly for testing and as a base experimenting
with the node affine extensions.
Michael, you could just take the second patch for initial load balancing
of your approach. Compared to your patch this one tries to achieve
equal node load also in the find_busiest_queue step. With more effort
and code, though... I hope it works for you, I removed the arch
dependencies and just need __cpu_to_node().
Regards,
Erich
-
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 : Mon Oct 07 2002 - 22:01:00 EST