cloudstack-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Darren Shepherd <darren.s.sheph...@gmail.com>
Subject Re: Whats involved in adding an extra hypervisor
Date Tue, 27 Aug 2013 15:21:57 GMT
Just to throw in my two cents.  I personally think the effort of adding a new hypervisor to
CloudStack is too difficult.  This is an area I'd like to focus on, along with another 15
things :)

I personally see a motivation for an xl integration.  This is something I've researched in
the past (and did).  Xen is very small, especially if you focus on only PV (with no qemu).
 You can build a Xen busybox system that is less that 50mb.  This opens up a great new way
to look at hypervisors in that you can PXE boot them and the whole OS is stateless and in
memory.  Basically what CoreOS is doing.  

Supporting a pure xl implementation is far more advanced thing though.  XenServer handles
a lot of nuisances of Xen, plus it makes storage easier (but networking more difficult). 
So if it was easy to add hypervisors I could see someone creating a specialized xl integration.
 If you want to support all the various networking and storage stuff, then libvirt or XenServer
are a better way to go. 

Darren

On Aug 27, 2013, at 7:10 AM, Sebastien Goasguen <runseb@gmail.com> wrote:

> 
> On Aug 19, 2013, at 3:09 PM, Chiradeep Vittal <Chiradeep.Vittal@citrix.com> wrote:
> 
>> On XenServer (from Citrix), there is no JVM, not sure if you can install
>> it even after the open sourcing.
> 
> I'd be surprised if you could not do it on Xen Project.
> 
>> On 8/19/13 10:38 AM, "Sebastien Goasguen" <runseb@gmail.com> wrote:
>> 
>>> 
>>> On Aug 19, 2013, at 1:25 PM, Chiradeep Vittal
>>> <Chiradeep.Vittal@citrix.com> wrote:
>>> 
>>>> I thought libvirt supports xen as well? Why "modify libvirt calls to
>>>> talk
>>>> to xl" ?
>>> 
>>> what I meant is that I believe our current Xen support makes use of xapi
>>> not libvrit.
>>> 
>>> So I would think we could "copy" the KVM agent and modify the "libvirt
>>> calls" -not libvrt itself- to use the xl tool.
>>> 
>>> That said I have not looked at the code for our Xen support.
>>> 
>>> 
>>>> On 8/19/13 6:40 AM, "Sebastien Goasguen" <runseb@gmail.com> wrote:
>>>> 
>>>>> Kind of on that same vein:
>>>>> 
>>>>> Anyone interested in modifying the KVM agent to support "pure" Xen
>>>>> without xapi.
>>>>> 
>>>>> I think it might be possible to just use the kvm agent and modify the
>>>>> libvirt calls to talk to xl , that way there would be no need for
>>>>> xapi...
>>>>> 
>>>>> Which if I am not mistaken would give us Xen on ARM support instantly
>>>>> in
>>>>> CloudStack
>>>>> 
>>>>> Any takers ? or anyone telling me I got this completely wrong ?
>>>>> 
>>>>> -Sebastien
>>>>> 
>>>>> On Aug 16, 2013, at 2:42 AM, Likitha Shetty <likitha.shetty@citrix.com>
>>>>> wrote:
>>>>> 
>>>>>> Ian, with AWS it¹s the other way around. The package allows us to
spin
>>>>>> up VMs in CS using AWS EC2 API's.
>>>>>> 
>>>>>> -Likitha
>>>>>> 
>>>>>>> -----Original Message-----
>>>>>>> From: Ian Duffy [mailto:ian@ianduffy.ie]
>>>>>>> Sent: Friday, August 16, 2013 12:07 PM
>>>>>>> To: CloudStack Dev
>>>>>>> Subject: Re: Whats involved in adding an extra hypervisor
>>>>>>> 
>>>>>>> Hi Donal and Chiradeep,
>>>>>>> 
>>>>>>> Thanks for your comments. It was an interesting read.
>>>>>>> 
>>>>>>> I might be missing something here but I will ask anyways. If
I
>>>>>>> understand
>>>>>>> correctly, at current with awsapi we are able to submit our aws
api
>>>>>>> credentials
>>>>>>> to Cloudstack and spin up VMs on aws correct? Is there a reason
the
>>>>>>> communication with aws could not be provided like a standard
>>>>>>> hypervisor? what
>>>>>>> is the reasoning behind keeping it as an almost separate package?
>>>>>>> 
>>>>>>> The reason I'm asking is because I'm wanting to do something
>>>>>>> Cloudstack based
>>>>>>> for a college project next year. However there is a hard 1 month
>>>>>>> deadline. I was
>>>>>>> interested to see could a base(or something of demoable quality)
for
>>>>>>> supporting
>>>>>>> Google Compute Engine be completed in such a short deadline.
>>>>>>> 
>>>>>>> On 15 August 2013 09:29, Donal Lafferty <donal.lafferty@citrix.com>
>>>>>>> wrote:
>>>>>>>> Definitely possible to add new Hypervisor types, if that's
what
>>>>>>>> you're asking.
>>>>>>>> 
>>>>>>>> How easy it is depends on how much existing CloudStack
>>>>>>>> infrastructure
>>>>>>>> you
>>>>>>> can exploit.  Let me out line the task with the help of some
planning
>>>>>>> questions:
>>>>>>>> 
>>>>>>>> 1. What will be your agent model?  Will you talk directly
to the
>>>>>>>> hypervisor
>>>>>>> (direct connect agent), or will you install a remote agent on
the
>>>>>>> hypervisor
>>>>>>> (connected agent).  This comes down to whether the hypervisor
exposes
>>>>>>> a high
>>>>>>> level API remotely.
>>>>>>>> 
>>>>>>>> 2. What will be your secondary storage model?  Secondary
storage
>>>>>>>> provides
>>>>>>> low IOPS storage accessible to all hypervisors in the zone. 
Thus, we
>>>>>>> store the
>>>>>>> templates in secondary storage.  IIRC, CloudStack supports NFS,
S3
>>>>>>> and
>>>>>>> Swift.
>>>>>>> Does one of these options suit your data centre, or do you need
to
>>>>>>> expand the
>>>>>>> list?  Will your agent be able to download disk images in secondary
>>>>>>> storage to
>>>>>>> the hypervisor?
>>>>>>>> 
>>>>>>>> 3. What will be your primary storage model?  Typically, primary
>>>>>>>> storage is
>>>>>>> high IOPS storage specific to a hypervisor or cluster of hypervisors.
>>>>>>> The easiest
>>>>>>> to setup is local storage, which can be a hard disk or storage
you
>>>>>>> mount
>>>>>>> manually on the hypervisor.  Alternatively, you may want to automate
>>>>>>> mounting
>>>>>>> storage on the hypervisor.
>>>>>>>> 
>>>>>>>> 4.  What will be your system VM model?  System VMs offload
the
>>>>>>>> following
>>>>>>> functionality from the management server:  VM console access,
>>>>>>> networking
>>>>>>> services, and secondary storage (upload) service.  You could
skip
>>>>>>> system VMs
>>>>>>> and run these services in the management server's host using
>>>>>>> QuickCloud.  You
>>>>>>> could run system VMs on an existing hypervisor type, or you could
add
>>>>>>> a new
>>>>>>> system VM type for your hypervisor.  Keep in mind that QuickCloud
>>>>>>> can't run
>>>>>>> your networking services.  Also, if you want to create a new
system
>>>>>>> VM
>>>>>>> type, you
>>>>>>> have to come up with VM image.
>>>>>>>> 
>>>>>>>> The tricky bits:
>>>>>>>> 
>>>>>>>> 5. What language will your agent use?  A direct connect agent
sits
>>>>>>>> in
>>>>>>>> the
>>>>>>> CloudStack process, so it is written in Java.  Alternatively,
there
>>>>>>> is
>>>>>>> infrastructure
>>>>>>> for a Java-based remote agent, which handles all your communications.
>>>>>>> If you
>>>>>>> need a non-Java remote agent, you are better off sending the
kernel
>>>>>>> commands
>>>>>>> over HTTP, which looks more like an RPC mechanism than REST.
>>>>>>>> 
>>>>>>>> 6. How will you know what instructions to implement?  You
can look
>>>>>>>> at
>>>>>>>> an
>>>>>>> existing ServerResource class for a hypervisor to know what command
>>>>>>> types
>>>>>>> there are.  The relevant pieces of data in each command can be
found
>>>>>>> out from
>>>>>>> existing hypervisor implementations.  Alternatively, you can
look at
>>>>>>> test logs,
>>>>>>> which contain the data from each command.  Eventually you'll
want to
>>>>>>> try your
>>>>>>> plugin with CloudStack itself.
>>>>>>>> 
>>>>>>>> Chiradeep's comments relate to #6 above.
>>>>>>>> 
>>>>>>>> 
>>>>>>>>> -----Original Message-----
>>>>>>>>> From: Chiradeep Vittal [mailto:Chiradeep.Vittal@citrix.com]
>>>>>>>>> Sent: 15 August 2013 02:51
>>>>>>>>> To: dev@cloudstack.apache.org
>>>>>>>>> Subject: Re: Whats involved in adding an extra hypervisor
>>>>>>>>> 
>>>>>>>>> Yes, it is a hypervisor plugin. While the extension method
may be
>>>>>>>>> simple, the impedance mismatch between the CloudStack
virtual model
>>>>>>>>> and the hypervisor API is what causes the most pain.
>>>>>>>>> E.g., CloudStack will hand a VirtualMachineTO object
(consisting of
>>>>>>>>> cpu/mem/nic) and then you have to use the hypervisor
API to
>>>>>>>>> construct it.
>>>>>>>>> For XS, it involves calling a bunch of XS APIs to 'construct'
the
>>>>>>>>> VM.
>>>>>>>>> For KVM, it involves constructing an XML file and passing
it to
>>>>>>>>> libvirt, etc.
>>>>>>>>> I'd say stuff like snapshots, stuff that involves a lot
of firewall
>>>>>>>>> configuration tends to be harder.
>>>>>>>>> 
>>>>>>>>> On 8/14/13 3:28 PM, "Ian Duffy" <ian@ianduffy.ie>
wrote:
>>>>>>>>> 
>>>>>>>>>> Hi Guys,
>>>>>>>>>> 
>>>>>>>>>> Just asking this off the top of my head with no research
done at
>>>>>>>>>> all.
>>>>>>>>>> Its a pure "Just out of interest" query.
>>>>>>>>>> 
>>>>>>>>>> Would it be a difficult task to add an interface
to Cloudstack in
>>>>>>>>>> order to enable it to communicate with some REST
based API that
>>>>>>>>>> goes
>>>>>>>>>> back to some hypervisor?
>>>>>>>>>> 
>>>>>>>>>> Can anybody point in the direction of code/files
I should look at
>>>>>>>>>> to
>>>>>>>>>> get an idea of the amount of work involved? Is the
plugin model in
>>>>>>>>>> such a state where such functionality could be abstracted
out as a
>>>>>>>>>> plugin?
>>>>>>>>>> With my previous experiences of dealing with the
cloudstack code
>>>>>>>>>> base I recall seeing a hypervisor folder in the plugins
folder, is
>>>>>>>>>> it just as simple as extending a few classes in there?
>>>>>>>>>> 
>>>>>>>>>> Thanks,
>>>>>>>>>> 
>>>>>>>>>> Ian
> 

Mime
View raw message