HI Dan:
Appreciate for
your presentation in summarizing the memory overcommit, really vivid and in
great help.
Well, I guess
recently days the strategy in my mind will fall into the solution Set C in pdf.
The tmem
solution your worked out for memory overcommit is both efficient and effective.
I guess I will
have a try on Linux Guest.
The real situation
I have is most of the running VMs on host are windows. So I had to come up
those policies to balance the memory.
Although
policies are all workload dependent. Good news is host workload is configurable,
and not very heavy
So I will
try to figure out some favorable policy. The policies referred in pdf are good
start for me.
Today, instead
of trying to implement “/proc/meminfo” with shared pages, I hacked the balloon
driver to have another
workqueue periodically
write meminfo into xenstore through xenbus, which solve the problem of xenstrore
high CPU
utilization problem.
Later I will
try to google more on how Citrix does.
Thanks for your
help, or do you have any better idea for windows guest?
Sent: Dan Magenheimer
[mailto:dan.magenheimer@xxxxxxxxxx]
Date: 2010.11.23 1:47
To: MaoXiaoyun; xen devel
CC: george.dunlap@xxxxxxxxxxxxx
Subject: RE: Xen balloon driver discuss
Xenstore IS slow and you could improve xenballoond performance
by only sending the single CommittedAS value from xenballoond in domU to dom0
instead of all of /proc/meminfo. But you are making an assumption
that getting memory utilization information from domU to dom0 FASTER (e.g. with
a shared page) will provide better ballooning results. I have not found
this to be the case, which is what led to my investigation into
self-ballooning, which led to Transcendent Memory. See the 2010 Xen Summit
for more information.
In your last paragraph below “Regards balloon strategy”, the
problem is it is not easy to define “enough memory” and “shortage of memory”
within any guest and almost impossible to define it and effectively load
balance across many guests. See my Linux Plumber’s Conference
presentation (with complete speaker notes) here:
http://oss.oracle.com/projects/tmem/dist/documentation/presentations/MemMgmtVirtEnv-LPC2010-Final.pdf
http://oss.oracle.com/projects/tmem/dist/documentation/presentations/MemMgmtVirtEnv-LPC2010-SpkNotes.pdf
From: MaoXiaoyun [mailto:tinnycloud@xxxxxxxxxxx]
Sent: Sunday, November 21, 2010 9:33 PM
To: xen devel
Cc: Dan Magenheimer; george.dunlap@xxxxxxxxxxxxx
Subject: RE: Xen balloon driver discuss
Since currently /cpu/meminfo is sent to domain 0 via xenstore, which in my opinoin is slow.
What I want to do is: there
is a shared page between domU and dom0, and domU
periodically
update the meminfo into the page, while on the other side dom0 retrive the
updated data for
caculating the target,
which is used by guest for balloning.
The problem I met is, currently I
don't know how to implement a shared page between
dom0 and domU.
Would it like dom 0 alloc a unbound event and wait guest to connect, and transfer date through
grant table?
Or someone has more efficient
way?
many thanks.
> From: tinnycloud@xxxxxxxxxxx
> To: xen-devel@xxxxxxxxxxxxxxxxxxx
> CC: dan.magenheimer@xxxxxxxxxx; George.Dunlap@xxxxxxxxxxxxx
> Subject: Xen balloon driver discuss
> Date: Sun, 21 Nov 2010 14:26:01 +0800
>
> Hi:
> Greeting first.
>
> I was trying to run about 24 HVMS (currently only Linux, later will
> involve Windows) on one physical server with 24GB memory, 16CPUs.
> Each VM is configured with 2GB memory, and I reserved 8GB memory for
> dom0.
> For safety reason, only domain U's memory is allowed to balloon.
>
> Inside domain U, I used xenballooned provide by xensource,
> periodically write /proc/meminfo into xenstore in dom
> 0(/local/domain/did/memory/meminfo).
> And in domain 0, I wrote a python script to read the meminfo, like
> xen provided strategy, use Committed_AS to calculate the domain U balloon
> target.
> The time interval is ! 1 seconds.
>
> Inside each VM, I setup a apache server for test. Well, I'd
> like to say the result is not so good.
> It appears that too much read/write on xenstore, when I give some of
> the stress(by using ab) to guest domains,
> the CPU usage of xenstore is up to 100%. Thus the monitor running in
> dom0 also response quite slowly.
> Also, in ab test, the Committed_AS grows very fast, reach to maxmem
> in short time, but in fact the only a small amount
> of memory guest really need, so I guess there should be some more to
> be taken into consideration for ballooning.
>
> For xenstore issue, I first plan to wrote a C program inside domain
> U to replace xenballoond to see whether the situation
> will be refined. If not, how about set up event channel directly for
> domU and dom0, would it be faster?
>
> Regards balloon strategy, I would do like this, when there ! are
> enough memory , just fulfill the guest balloon request, and when shortage
> of memory, distribute memory evenly on the guests those request
> inflation.
>
> Does anyone have better suggestion, thanks in advance.
>
