From mboxrd@z Thu Jan 1 00:00:00 1970 From: Robin Holt Date: Tue, 15 Feb 2005 20:08:56 +0000 Subject: Re: [PATCH] Convert pgtable cache to slab Message-Id: <20050215200856.GG24401@lnx-holt.americas.sgi.com> List-Id: References: In-Reply-To: MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit To: linux-ia64@vger.kernel.org On Tue, Feb 15, 2005 at 12:03:22PM -0800, David Mosberger wrote: > >>>>> On Tue, 15 Feb 2005 13:59:45 -0600, Robin Holt said: > > >> I certainly don't want to switch my machines to 4-levels. There > >> is zero need for that on the machines I use, so why pay the > >> overhead of an extra level? > > Robin> I missed the why the extra level question. That part is > Robin> motivated by some large MPI jobs need a fourth level to get a > Robin> large enough single mapping to cover their entire dataset. > Robin> We have already tripped this limitation for a couple of our > Robin> really large customers. > > What I'm saying is that the default probably should stay at 3-levels. > Optimizing ia64 linux _just_ for "really large customers" would be > a bad direction to go in. I agree. What I am asking is if we turn on the 4th level of page tables, does anybody have an objection to 4/16k levels as opposed to 3/PAGE_SIZE plus a 4th cleanup to cover the entire address space? Thanks, Robin