On Tue, Apr 22, 2014 at 08:34:53PM +0700, Sergiusz Pawlowicz wrote:
> On Tue, Apr 22, 2014 at 8:21 PM, Herbert Poetzl <email@example.com> wrote:
>> On Tue, Apr 22, 2014 at 02:15:27PM +0700, Sergiusz Pawlowicz wrote:
>>> On Tue, Apr 22, 2014 at 2:05 PM, Urban Loesch <firstname.lastname@example.org> wrote:
>>>> For me the only solution now is to give the guest enough
>>>> memory so that it can't reach the limits.
>>>> From my experience, on 3.4x line cgroups are able to enforce limits,
>>> on later kernels whole server hangs. But sticking to 3.4x is
>>> not a long term solution.
>>> If the issue won't get resolved, unfortunately it means we will
>>> have to escape from linux-vserver to KVM, where cgroups work
>>> properly :-(
>> We do not modify the mainline cgroup system, so it is the same
>> system whether you use it for Linux-VServer or KVM, which in
>> turn means, if it is broken in one of the upstream kernels
>> (and it looks like it has some issues in certain kernel versions)
>> it will be broken for Linux-VServer (and KVM) as well.
> Hi Herbert,
> no, KVM, which I have tested, is not taking down the whole
> server. It OOMs and kills the process which is taking the
I might not have explained it well, but facts are:
* we do not modify how cgroups work in Linux-VServer
-> so they will work like on any other Linux system
-> they will be as good or as bad as on any other system
-> they will do the same in LXC or Linux-VServer
* cgroups are the same for host and guests
-> protecting the host is your obligation
-> configuring the guest limits is unrelated to Linux-VServer
> I'm not sure if anything is "broken", it seems vserver way of
> implementing cgroup limits is not intelligent enough.
Something you probably have to get sorted upstream/mainline.
> From my understanding, and previous attempts to explain the
> issue I could read on the mailing list, if vserver guest is
> OOMing, cgroup enforces complete blockade not only on the RAM,
> but also of the filesystem, which in consequence blocks the
> whole server, which hangs.
> It is major issue, which must be investigated (and is easy to
> simulate) or vserver technology will be outdated.
Same is true for LXC, and I doubt that mainline will not care
but as I said, you are barking up the wrong tree, we used to
use our own memory accounting and limits which worked quite
fine with occasional hiccups, but we switched to the mainline
cgroup system some time ago because that is how we roll (i.e.
incorporating upstream as soon as it becomes somewhat useable)
I'm pretty sure you will see the same issues with and even
in KVM, but as KVM is completely different (separate kernel,
separate resources) it won't give the same results if you
compare an LXC or Linux-VServer guest with a KVM one.
Received on Tue Apr 22 15:03:07 2014