From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1762254Ab3JQDoQ (ORCPT ); Wed, 16 Oct 2013 23:44:16 -0400 Received: from mail-pd0-f170.google.com ([209.85.192.170]:54880 "EHLO mail-pd0-f170.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1761650Ab3JQDoO (ORCPT ); Wed, 16 Oct 2013 23:44:14 -0400 Date: Wed, 16 Oct 2013 20:44:12 -0700 (PDT) From: David Rientjes X-X-Sender: rientjes@chino.kir.corp.google.com To: Anton Vorontsov cc: Andrew Morton , Michal Hocko , "Kirill A. Shutemov" , linux-kernel@vger.kernel.org, linux-mm@kvack.org Subject: Re: [patch] mm, vmpressure: add high level In-Reply-To: <20131017030512.GA21327@teo> Message-ID: References: <20131017030512.GA21327@teo> User-Agent: Alpine 2.02 (DEB 1266 2009-07-14) MIME-Version: 1.0 Content-Type: TEXT/PLAIN; charset=US-ASCII Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Wed, 16 Oct 2013, Anton Vorontsov wrote: > > Vmpressure has two important levels: medium and critical. Medium is > > defined at 60% and critical is defined at 95%. > > > > We have a customer who needs a notification at a higher level than medium, > > which is slight to moderate reclaim activity, and before critical to start > > throttling incoming requests to save memory and avoid oom. > > > > This patch adds the missing link: a high level defined at 80%. > > > > In the future, it would probably be better to allow the user to specify an > > integer ratio for the notification rather than relying on arbitrarily > > specified levels. > > Does the customer need to differentiate the two levels (medium and high), > or the customer only interested in this (80%) specific level? > Only high. > In the latter case, instead of adding a new level I would vote for adding > a [sysfs] knob for modifying medium level's threshold. > Hmm, doesn't seem like such a good idea. If one process depends on this being 60% and another depends on it being 80%, we're stuck. I think it's legitimate to have things like low, medium, high, and critical as rough approximations (and to keep backwards compatibility), but as mentioned in the changelog I want to extend the interface to allow integer writes to specify their own ratio.