Re: [Xen-devel] can not use all available memory

[Sander Eikelenboom <linux@xxxxxxxxxxxxxx>]

Monday, November 26, 2012, 9:20:40 PM, you wrote:

>> From: Sander Eikelenboom [mailto:linux@xxxxxxxxxxxxxx]
>> Subject: Re: [Xen-devel] can not use all available memory
>> 
>> Monday, November 26, 2012, 5:58:28 PM, you wrote:
>> 
>> >> From: Ian Campbell [mailto:Ian.Campbell@xxxxxxxxxx]
>> >> Sent: Friday, November 23, 2012 6:34 AM
>> >> To: Pasi KÃrkkÃinen
>> >> Cc: xen-devel@xxxxxxxxxxxxxxxxxxx; Alexander Bienzeisler
>> >> Subject: Re: [Xen-devel] can not use all available memory
>> >>
>> >> On Fri, 2012-11-23 at 13:29 +0000, Pasi KÃrkkÃinen wrote:
>> >> > On Fri, Nov 23, 2012 at 01:15:53PM +0000, Ian Campbell wrote:
>> >> > > On Fri, 2012-11-23 at 12:55 +0000, Pasi KÃrkkÃinen wrote:
>> >> > >
>> >> > > > > This has been discussed at length on the list before, please 
>> >> > > > > check the
>> >> > > > > archives.
>> >> > > > >
>> >> > > >
>> >> > > > I tried googling quickly but I didn't find anything relevant..
>> >> > >
>> >> > > I found
>> >> > > http://web.archiveorange.com/archive/v/zCKz5T3PLvtyZDSPQc9i
>> >> > > in a matter of seconds, then:
>> >> > > http://lists.xen.org/archives/html/xen-devel/2011-11/msg01415.html
>> >> > > http://lists.xen.org/archives/html/xen-users/2012-05/msg00139.html
>> >> > > http://lists.xen.org/archives/html/xen-users/2012-05/msg00146.html
>> >> > >
>> >> > > > To me this behaviour still seems wrong. What's the point of 
>> >> > > > autoballoon=1 trying to
>> >> > > > balloon down dom0 if the hypervisor already has enough free memory 
>> >> > > > for the VM ?
>> >> > > >
>> >> > > > In this case:
>> >> > > >         - dom0_mem=2G
>> >> > > >         - new VM to launch with size 16 GB.
>> >> > > >         - Xen has 28 GB of free memory.
>> >> > > >
>> >> > > > So clearly there's no need to try to balloon down dom0..
>> >> > >
>> >> > > Right, so don't set autoballoon then.
>> >> > >
>> >> >
>> >> > It's enabled as a default.. so many people hit this problem.
>> >> >
>> >> >
>> >> > > > not-yet-implemented check/feature in xl, or a bug?
>> >> > >
>> >> > > Neither, it is the intended behaviour of xl autoballoon, this option
>> >> > > means exactly "take the required memory from dom0".
>> >> > >
>> >> >
>> >> > http://xenbits.xen.org/docs/4.2-testing/man/xl.conf.5.html
>> >> >
>> >> > "autoballoon=BOOLEAN
>> >> >
>> >> >     If disabled then xl will not attempt to reduce the amount of
>> >> > memory assigned to domain 0 in order to create free memory when
>> >> > starting a new domain. You should set this if you use the dom0_mem
>> >> > hypervisor command line to reduce the amount of memory given to domain
>> >> > 0 by default.
>> >> >
>> >> >     Default: 1"
>> >> >
>> >> >
>> >> > I think we should modify that to say "You should set autoballoon=0 if 
>> >> > you use the dom0_mem
>> >> hypervisor command line .."
>> >> > At least I understood that text in the opposite way..
>> >>
>> >> Yes, there should be a s/set/unset/ in there I think.
>> >>
>> >> > Also: What happens if you have autoballoon=1 and you start some VMs, 
>> >> > then stop the VMs,
>> >> > so most of the memory is now free in Xen..
>> >>
>> >> xl balloons dom0 back up when it destroy domains with autoballoon=1.
>> >>
>> >> > and then you try to start a big VM ?
>> >> > Aren't you going to hit the same problem as in this thread?
>> 
>> > Hmmm... this behavior and default may make sense for the Citrix
>> > memory model (single machine, dom0 is "the user" so you want it
>> > to always hold most of physical RAM not used by guests).  But
>> > probably not so for a more cloud-like memory model.
>> 
>> > Is there any (easy) way to force autoballoon=0 if the hypervisor
>> > dom0_mem boot option is specified?  Or is there some reasonably
>> > sane case I am missing where a user would want both dom0_mem
>> > and autoballoon=1?
>> 
>> > Oracle VM always boots servers with dom0_mem= set so (if/when
>> > OVM switches to use xl), OVM will always set autoballoon off.
>> > So it's the large number of non-Citrix-non-Oracle Xen-as-a-service
>> > providers I am trying to help here.
>> 
>> 
>> 
>> Hmm i was bitten by this 2 weeks ago, i found it a bit not intuitive that:
>> - While the guest i was trying to start required less memory than was freely 
>> available (according to
>> xentop) outside of dom0
>> - It would fail, because xl started to try to balloon dom0 down which failed.

> Hi Sander --

> I could be wrong (and I am confident someone will correct me if I am) but
> I think this is because the Citrix memory model assumes there is an
> inference-driven policy engine for load-balancing memory across competing
> virtual machines ("squeezed").  I suspect squeezed returns unallocated
> xen "free" memory to dom0.

> IMHO, such policy engines are good for demos and so salespeople can
> say "yes, this product supports memory overcommit" but Transcendent
> Memory goes quite a bit further (albeit not for guests with proprietary
> OS's).

> Dan

Yes well my novice assumption was that if i have say 8GB of mem, and set dom0 
max mem to say 1GB.
That auto ballooning would not interfere unless the total of guests i start go 
across the 7GB.

And it seems when i go under the, say about 1GB of free mem available outside 
of dom0, some ballooning code seems to interfere with creating guests.
Most of the time, the first attempt to start a guests then fails, a second 
attempt directly thereafter succeeds.

Switching off autoballoon in xl.conf makes it all work fine. Not a big problem 
in itself, but as Pasi said perhaps not quite clear and intuitive for everyone.

So your suggestion of disabling autoballooning when dom0_mem is set seems 
appealing. Although there perhaps is some use case in limiting dom0 to under 
4GB for some drivers, but still leave the capability to autoballoon down ?

(and yes i don't like all the automatic voodoo in overcommiting memory and 
"hope for the best" all the algorithms involved can manage it.
I like things to predictably fail (and fail direct if possible), so good 
measures can be taken. A bit old-fashioned perhaps.

--
Sander





_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel