----- Original Message -----
On Fri, 07 Sep 2012 13:27:15 +0530, "M. Mohan Kumar" mohan@in.ibm.com wrote:
On Fri, 07 Sep 2012 09:35:10 +0300, Itamar Heim iheim@redhat.com wrote:
On 09/07/2012 08:21 AM, M. Mohan Kumar wrote:
On Thu, 6 Sep 2012 18:59:19 -0400 (EDT), Ayal Baron abaron@redhat.com wrote:
For start using the LVs we will always do truncate for the required size, it will resize the LV. I didn't get what you are mentioning about thin-provisioning, but I have a dumb code using dm-thin targets showing BD xlators can be extended to use dm-thin targets for thin-provisioning.
so even though this is block storage, it will be extended as needed? how does that work exactly? say i have a VM with a 100GB disk. thin provisioning means we only allocated 1GB to it, then as the guest uses that storage, we allocate more as needed (lvextend, pause guest, lvrefresh, resume guest)
When we use device=lv, it means we use only thick provisioned logical volumes. If this logical volume runs out of space in the guest, one can resize it from the client by using truncate (results in lvresize at the server side) and run filesystem tools at guest to get added space.
But with device=thin type, all LVs are thinly provisioned and allocating space to them is taken care by device-mapper thin target automatically. The thin-pool should have enough space to accomoodate the sizing requirements.
As of now BD xlator supports only working with linear Logical volumes, they are thick provisioned. gluster cli command "gluster volume create" with option "device=lv" allows to work with logical volumes as files.
As a POC I have a code(not posted to external list), with option "device=thin" to gluster volume create command it allows to work with thin provisioned targets. But it does not take care of resizing thin-pool when it reaches low-level threshold. Supporting thin targets is in our TODO list. We have dependency on lvm2 library to provide apis to create thin-targets.
I'm definitely missing some background here. 1. Can the LV span on multiple bricks in gLuster? i. If 'yes' then a. do you use gLuster's replication and distribution schemes to gain performance and redundancy? b. what performance gain is there over normal gLuster with files? ii. If 'not' then you're only exposing single host local storage LVM? (in which case I don't see why gLuster is used at all and where).
From a different angle, the only benefit I can think of in exposing a fs interface over LVM is for consumers who do not wish to know the details of the underlying storage but want the performance gain of using block storage.
vdsm is already intimately familiar with LVM and block devices, so adding the FS layer scheme on top doesn't strike me as adding any value. In addition, you require the consumer to know a lot about your interface because it's not truely a FS interface. e.g. consumer is not allowed to create directories, files are not sparse, not to mention that if you're indeed using LVM then I don't think you're considering the VG MD and extent size limitations: 1. LVM currently has severe limitations wrt number of objects it can manage (the limitation is actually the size of the VG metadata, but the distinction is not important just yet). This means that creating a metadata LV in addition to each data LV is very costly (at around 1000 LVs you'd hit a problem. vdsm currently creates 2 files per snapshot (the data and a small file with metadata describing it) meaning that you'd reach this limit really fast. 2. LVM max LV size is extent size * 65K, this means that if I choose a 4K extent size then my max LV size would be 256MB. This obviously won't do for VMs disks so you'd choose a much larget extent size. However a larger extent size means that each metadata file vdsm creates wastes a lot of storage space. So even if LVM could scale, your storage usage plummets and your $/MB ratio increases. The way around this is of course not to have a metadata file per volume but have 1 file containing all the metadata, but then that means I'm fully aware of the limitations of the environment and treating my objects as files gains me nothing (but does require a new hybrid domain, a lot more code etc).
Also note that without thin provisioning we loose our ability to create snapshots.
vdsm-devel mailing list vdsm-devel@lists.fedorahosted.org https://lists.fedorahosted.org/mailman/listinfo/vdsm-devel