From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: from mail-it0-f50.google.com ([209.85.214.50]:55682 "EHLO mail-it0-f50.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751517AbdJLPz4 (ORCPT ); Thu, 12 Oct 2017 11:55:56 -0400 Received: by mail-it0-f50.google.com with SMTP id l196so7379354itl.4 for ; Thu, 12 Oct 2017 08:55:55 -0700 (PDT) Subject: Re: btrfs seed question To: Chris Murphy , Joseph Dunn Cc: Anand Jain , Btrfs BTRFS References: <20171011204759.1848abd7@olive.ig.local> <20171012092028.1fbe79d9@olive.ig.local> <20171012104438.4304e99a@olive.ig.local> From: "Austin S. Hemmelgarn" Message-ID: <13ec83e8-f0a1-0f55-b755-00a02d68791d@gmail.com> Date: Thu, 12 Oct 2017 11:55:51 -0400 MIME-Version: 1.0 In-Reply-To: Content-Type: text/plain; charset=utf-8; format=flowed Sender: linux-btrfs-owner@vger.kernel.org List-ID: On 2017-10-12 11:30, Chris Murphy wrote: > On Thu, Oct 12, 2017 at 3:44 PM, Joseph Dunn wrote: >> On Thu, 12 Oct 2017 15:32:24 +0100 >> Chris Murphy wrote: >> >>> On Thu, Oct 12, 2017 at 2:20 PM, Joseph Dunn wrote: >>>> >>>> On Thu, 12 Oct 2017 12:18:01 +0800 >>>> Anand Jain wrote: >>>> >>>>> On 10/12/2017 08:47 AM, Joseph Dunn wrote: >>>>>> After seeing how btrfs seeds work I wondered if it was possible to push >>>>>> specific files from the seed to the rw device. I know that removing >>>>>> the seed device will flush all the contents over to the rw device, but >>>>>> what about flushing individual files on demand? >>>>>> >>>>>> I found that opening a file, reading the contents, seeking back to 0, >>>>>> and writing out the contents does what I want, but I was hoping for a >>>>>> bit less of a hack. >>>>>> >>>>>> Is there maybe an ioctl or something else that might trigger a similar >>>>>> action? >>>>> >>>>> You mean to say - seed-device delete to trigger copy of only the >>>>> specified or the modified files only, instead of whole of seed-device ? >>>>> What's the use case around this ? >>>>> >>>> >>>> Not quite. While the seed device is still connected I would like to >>>> force some files over to the rw device. The use case is basically a >>>> much slower link to a seed device holding significantly more data than >>>> we currently need. An example would be a slower iscsi link to the seed >>>> device and a local rw ssd. I would like fast access to a certain subset >>>> of files, likely larger than the memory cache will accommodate. If at >>>> a later time I want to discard the image as a whole I could unmount the >>>> file system or if I want a full local copy I could delete the >>>> seed-device to sync the fs. In the mean time I would have access to >>>> all the files, with some slower (iscsi) and some faster (ssd) and the >>>> ability to pick which ones are in the faster group at the cost of one >>>> content transfer. >>> >>> >>> Multiple seeds? >>> >>> Seed A has everything, is remote. Create sprout B also remotely, >>> deleting the things you don't absolutely need, then make it a seed. >>> Now via iSCSI you can mount both A and B seeds. Add local rw sprout C >>> to seed B, then delete B to move files to fast local storage. >>> >> Interesting thought. I haven't tried working with multiple seeds but >> I'll see what that can do. I will say that this approach would require >> more pre-planning meaning that the choice of fast files could not be >> made based on current access patterns to tasks at hand. This might >> make sense for a core set of files, but it doesn't quite solve the >> whole problem. > > > I think the use case really dictates a dynamic solution that's smarter > than either of the proposed ideas (mine or yours). Basically you want > something that recognizes slow vs fast storage, and intelligently > populates fast storage with frequently used files. > > Ostensibly this is the realm of dmcache. But I can't tell you whether > dmcache or via LVM tools, if it's possible to set the proper policy to > make it work for your use case. And also I have no idea how to set it > up after the fact, on an already created file system, rather than > block devices. It is possible with dm-cache, but not at the file level (as befits any good block layer, it doesn't care about files, just blocks). Despite that, it should work reasonably fine (I've done this before with NBD and ATAoE devices, and it worked perfectly, so I would expect it to work just as well for iSCSI), and actually may do better than caching whole files locally depending on your workload. As far as setup after the fact, the degree of difficulty depends on whether or not you want to use LVM. Without LVM, you should have no issue just setting up a device-mapper table for caching, you just need enough room on the local SSD for both the cache data and cache metadata partition. When you create the table using dmsetup 9and eventually in /etc/dmtab), it won't wipe the filesystem on the origin device, so you can easily add a cache to anything this way, even if it's an existing filesystem (though you will need to unmount the filesystem to add the cache). All things considered, it's no worse than setting it up on a brand new device, the hardest part is making sure you get the device sizes right for the device mapper table. With LVM however, it's somewhat more complicated, because it refuses to work with anything it's not managing already, so you would have to reprovision the iSCSI device as a PV, add it to the local VG, and then work with that. If you don't mind reprovisioning, I would actually suggest bcache instead of LVM here though. It's less complicated to add and remove caches, does somewhat better in recent versions of intelligently deciding what to cache locally, and is also significantly less likely to slow down the rest of your system than LVM (any management operations will have to hit that remote iSCSI device, and LVM does a lot more with data on the disk than bcache does). > > The hot vs cold files thing, is something I thought the VFS folks were > looking into. I was under that impression too, but I haven't seen anything relating to it recently (though I'm not subscribed to the linux-vfs list, so there may be discussion there I'm not seeing).