It's consistent to call node_online() with a physical node ID when the
online node mask is composed of fake nodes?
I don't understand your comments. My code is intended to work for all
systems. If the system is non-NUMA by nature, then all CPUs map to fake
node 0.
As an example, on a two chip dual-core AMD opteron system, there are 4
"cpus" where CPUs 0 and 1 are close to the first half of memory, and
CPUs 2 and 3 are close to the second half. Without this change CPUs 2
and 3 are mapped to fake node 1. This results in awful performance. With
this change, CPUs 2 and 3 are mapped to (roughly) 1/2 the fake node
count. Their zonelists[] are ordered to do allocations preferentially
from zones that are local to CPUs 2 and 3.
Can you tell me the scenario where my code makes things worse?
I'm adding no data structures to do this. The current code already has
get_phys_node. My changes use the existing information about node
layout, both the physical and fake, and defines a mapping. The current
mapping just takes a physical node and says "it's the fake node too".
I wish you would include some specifics as to why you think what you
do. You're suggesting we leave in place a system that destroys NUMA
locality when using fake numa, and passes around physical node ids as an
index into nodes[] whihc is indexed by fake nodes. My change has no
effect without fake numa, and harms no one with fake numa.
-- Ethan
-
