Messages in this thread | | | Date | Mon, 21 May 2012 10:40:46 +0200 | From | Ingo Molnar <> | Subject | Re: [tip:sched/numa] sched/numa: Introduce sys_numa_{t,m}bind() |
| |
* David Rientjes <rientjes@google.com> wrote:
> On Fri, 18 May 2012, tip-bot for Peter Zijlstra wrote: > > > Commit-ID: 3a0fea961b98d1838f35dba51a639d40f4a5589f > > Gitweb: http://git.kernel.org/tip/3a0fea961b98d1838f35dba51a639d40f4a5589f > > Author: Peter Zijlstra <a.p.zijlstra@chello.nl> > > AuthorDate: Thu, 8 Mar 2012 17:56:08 +0100 > > Committer: Ingo Molnar <mingo@kernel.org> > > CommitDate: Fri, 18 May 2012 08:16:27 +0200 > > > > sched/numa: Introduce sys_numa_{t,m}bind() > > > > This depends on 931ea9d1a6e0 ("rcu: Implement per-domain > single-threaded call_srcu() state machine") from core/rcu.
Indeed ...
I'll rebase it to a (by that time probably upstream) srcu commit after the merge window, once we have more fixes, have incorporated suggestions, etc. - but it's still essentially an RFC topic: [*]
Fundamentally, do people agree with the 'single home node' approach to begin with? We could turn it into a node mask, but that complicates things.
For example if there's a hierarchy of nodes, low latency and high latency ones, then it might be valid to limit to a high level (high latency) node and not specify the lower level node - while the kernel would still know about the lower level nodes as well.
Managing locality in a non-trivial cache hierarchy is hard :-/
Thanks,
Ingo
[*] I should probably move this to the tip:RFC/sched/numa branch, to make it clear what the status of the branch is, from the commit notification emails.
| |