Lazy NUMA sorting?

Lazy NUMA sorting?

[Timothy Miller]

So, I understand that the 2.6 SMP balancer redistributes workload on a 
periodic basis.  Once every second or something, it migrates processes.

A NUMA system would have to do something similar, where if there is a 
page which is referenced by only one process, and the page is located on 
the "wrong" node, it could be migrated.  This could be done gradually by 
a periodic background process.  Is this already how it works?

Also, if a page is being referenced by multiple nodes, the same 
background process could make mirror copies.  (Age of page would be an 
important consideration here so moves don't happen for short-lived pages.)

Re: Lazy NUMA sorting?

[Martin J. Bligh]

> So, I understand that the 2.6 SMP balancer redistributes workload on a periodic basis.  Once every second or something, it migrates processes.
> 
> A NUMA system would have to do something similar, where if there is a page which is referenced by only one process, and the page is located on the "wrong" node, it could be migrated.  This could be done gradually by a periodic background process.  Is this already how it works?

We don't do page migration. It has proved extraordinarily hard to know when
to migrate pages correctly ... in conjunction with the cross-node movement
of tasks.

> Also, if a page is being referenced by multiple nodes, the same background process could make mirror copies.  (Age of page would be an important consideration here so moves don't happen for short-lived pages.)

We can do that for read-only mappings and kernel text already. Dave Hansen
did a patch for the read-only mappings, though I need to update it to the
latest kernel tree.

M.