From mboxrd@z Thu Jan 1 00:00:00 1970 From: Arnd Bergmann Subject: Re: NUMA BOF @OLS Date: Fri, 22 Jun 2007 12:14:58 +0200 Message-ID: <200706221214.58823.arnd@arndb.de> References: <200706220112.51813.arnd@arndb.de> Mime-Version: 1.0 Content-Type: text/plain; charset=iso-8859-1 Content-Transfer-Encoding: QUOTED-PRINTABLE Return-path: In-Reply-To: Content-Disposition: inline Sender: linux-kernel-owner@vger.kernel.org To: Christoph Lameter Cc: linux-mm@kvack.org, linux-kernel@vger.kernel.org List-Id: linux-mm.kvack.org On Friday 22 June 2007, Christoph Lameter wrote: >=20 > On Fri, 22 Jun 2007, Arnd Bergmann wrote: >=20 > > - Interface for preallocating hugetlbfs pages per node instead of s= ystem wide >=20 > We may want to get a bit higher level than that. General way of=20 > controlling subsystem use on nodes. One wants to restrict the slab=20 > allocator and the kernel etc on nodes too. >=20 > How will this interact with the other NUMA policy specifications? I guess that's what I'd like to discuss at the BOF. I frequently get requests from users that need to have some interface for it: Application currently break if they try to use /proc/sys/vm/nr_hugepage= s in combination with numactl --membind. > > - architecture independent in-kernel API for enumerating CPU socket= s with > > =A0 multicore processors (not sure if that's the same as your exist= ing subject). >=20 > Not sure what you mean by this. We already have a topology interface = and=20 > the scheduler knows about these things. I'm not referring to user interfaces or scheduling. It's probably not r= eally a NUMA topic, but we currently use the topology interfaces for enumerat= ing sockets on systems that are not really NUMA. This includes stuff like per-socket=20 * cpufreq settings (these have their own logic currently) * IOMMU * performance counters * thermal management * local interrupt controller * PCI/HT host bridge If you have a system with multiple CPUs in one socket and either multip= le sockets in one NUMA node or no NUMA at all, you have no way of properl= y enumerating the sockets. I'd like to discuss what such an interface would need to look like to be useful for all architectures. Arnd <><