From: Frank Mayhar Subject: Re: [dm-devel] dm: Make MIN_IOS, et al, tunable via sysctl. Date: Tue, 20 Aug 2013 14:52:45 -0700 Message-ID: <1377035565.24869.53.camel@bobble.lax.corp.google.com> References: <1376070533.26057.244.camel@bobble.lax.corp.google.com> <20130819140016.GB27167@redhat.com> <1376934845.24869.43.camel@bobble.lax.corp.google.com> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: Sender: linux-kernel-owner@vger.kernel.org To: Mikulas Patocka Cc: device-mapper development , linux-kernel@vger.kernel.org List-Id: dm-devel.ids On Tue, 2013-08-20 at 17:44 -0400, Mikulas Patocka wrote: > On Mon, 19 Aug 2013, Frank Mayhar wrote: > > On Mon, 2013-08-19 at 10:00 -0400, Mike Snitzer wrote: > > > Performance isn't the concern. The concern is: does DM allow for > > > forward progress if the system's memory is completely exhausted? > > > > > > This is why request-based has such an extensive reserve, because it > > > needs to account for cloning the largest possible request that comes in > > > (with multiple bios). > > > > Thanks for the response. In our particular case, I/O will be file > > system based and over a network, which makes it pretty easy for us to be > > sure that large I/Os never happen. That notwithstanding, however, as > > you said it just seems reasonable to make these values configurable. > > > > I'm also looking at making some similar constants in dm-verity and > > dm-bufio configurable in the same way and for similar reasons. > > Regarding dm-bufio: the user of dm-bufio sets the pool size as an argument > in dm_bufio_client_create. There is no need to make it configurable - if > the user selects too low value, deadlock is possible, if the user selects > too high value, there is no additional advantage. True, but dm-bufio also allocates a a fixed-size 8MB (on a 64-bit machine) hash table. I'm still getting performance data but it appears that reducing this, even by a lot, doesn't impact performance significantly, at least not with the workload I'm running. (Which is using fio, random and sequential reads of varying buffer sizes.) > Regarding dm-verity: the mempool size is 4, there is no need to make it > bigger, there is no advantage from that. Also true, but there may be an advantage in making it smaller. -- Frank Mayhar 310-460-4042