Messages in this thread | | | Subject | Re: [Xen-devel] [PATCH v1][RFC] drivers/xen, balloon driver numa support in kernel | From | Dario Faggioli <> | Date | Mon, 12 Aug 2013 22:14:43 +0200 |
| |
On lun, 2013-08-12 at 19:44 +0100, David Vrabel wrote: > On 12/08/13 15:13, Yechen Li wrote: > > This small patch adds numa support for balloon driver. Kernel version: 3.11-rc5 > > It's just a RFC version, since I'm waiting for the interface of numa topology. > > The balloon driver will read arguments from xenstore: /local/domain/(id)/memory > > /target_nid, and settle the memory increase/decrease operation on specified > > p-nodeID. > > Its is difficult to review an ABI change without any documentation for > the new ABI. > Indeed.
> I would also like to see a design document explaining the overall > approach planned to be used here. It's not clear why explicitly > specifying nodes is preferable to (e.g.) the guest releasing/populating > evenly across all its nodes (this would certainly be better for the guest). > I see what you mean. Personally, I think they're different things. There might be the need, from the host system administrator, to make as much room as possible on one (or perhaps a few) nodes, in which case, the possibility of specifying that explicitly would be a plus.
That would allow --if used wisely, I agree with you on this-- for better resource utilization, in the long run.
In absence of this information, it is probably true that the guest would benefit from a more even approach. What we want to achieve here, however, is as follows: suppose that a virtual NUMA enabled guest (i.e., a guest with a virtual NUMA topology), has guest page X, which is on virtual node g1 in the guest itself, backed by a frame from host node h0. Well, we really would like to try having page X always backed by a frame on host node h1, even after ballooning down and up.
> It seems like unless this is used carefully, all VMs will end up with > suboptimal memory layouts as they are repeatedly balloon up and down to > satisfy the whims of the latest VM being started etc. > I'm not sure I see entirely what you mean, but for sure I repeat that I agree that more information about the design and intended usage patterns are needed... Let's see whether Yechen is up for providing that. :-)
Thanks for having a look anyway, Dario
-- <<This happens because I choose it to happen!>> (Raistlin Majere) ----------------------------------------------------------------- Dario Faggioli, Ph.D, http://about.me/dario.faggioli Senior Software Engineer, Citrix Systems R&D Ltd., Cambridge (UK)
[unhandled content-type:application/pgp-signature] | |