[Vm-dev] Re: [Pharo-dev] Cog in the cloud

[Alain Rastoul]

Le 22/12/2014 13:14, Ben Coman a écrit :
>
> Ben Coman wrote:
>>
>> (whoops forgot to cross-post)
>>
>> Eliot Miranda wrote:
>>> Hi Ben,
>>>
>>>
>>> On Dec 19, 2014, at 8:08 PM, Ben Coman <btc at openInWorld.com> wrote:
>>>
>>>> Just some thoughts that arose as I skipped along the web...
>>>>
>>>> With the idea of using Pharo in the cloud, I was thinking of how
>>>> PharoNOS[1] is basing off top of the Linux kernel, but the examples
>>>> indicate its operating in User Mode.  So I was wondering about
>>>> performance being greater[2] if the CogVM ran in Kernel Mode.
>>>> However maybe this could get tangled in the GPL license of the kernel.
>>>>
>>>> Now I believe FreeBSD/NetBSD license is compatible with COG's MIT
>>>> license, so that seems a better option for eliminating layers of the
>>>> operating system.  They both[3][4] have pre-built Amazon Machine
>>>> Images to run on EC2, as well as instructions[5] to build your own
>>>> AMI.  So potentially we could build an AMI with the CogVM linked to
>>>> the FreeBSDKernel operating in Kernel Mode, with no User Mode. The
>>>> lack of direct access to memory from the Image, and probably single
>>>> application focus should be sufficient security to forgo User Mode,
>>>> and run faster.
>>>>
>>>> But we could go a step further.  Runing in the cloud relies heavily
>>>> on virtualisation, and for performance[6], most likely on the OS's
>>>> paravirtualisation interface - to Xen for example.  So from Cog, why
>>>> not interface direct to the Xen FrontEndDriver[7] and eliminate the
>>>> DomU operating system all together.  What I understand from [6] is
>>>> that Xen's paravirtualisation hooks makes it much easier to boot,
>>>> than booting on the bare metal of a fully-virtualised system. So the
>>>> former seems more achievable than the latter (which was the only
>>>> option in years past), and also be more portable as technology
>>>> evolves, to any platform Xen runs on.
>>>>
>>>> Squeak/Pharo on Cog-Xen could be a good choice for a Cloud Operating
>>>> System[8]. Maybe a good student research project?
>>>
>>> This sounds really cool, but could you analyse a bit further?
>>>
>>> - what key economic and managerial benefits accrue from this
>>> configuration?
>>
>> I'll put some more thought into this, but just off the top of my head
>> (so it turned out more technical than economic/managerial):
>>
>> * Eliminates performance overhead of system call context switching.
>> Although this is replaced by hypercalls, it removes one layer on
>> hypervisor systems, which plausibly is going to be EVERYTHING in the
>> future.
>>
>> * Eliminate performance overhead of OS networking stack
>> http://roscidus.com/blog/blog/2014/08/15/optimising-the-unikernel/
>>
>> * Data from virtual devices is event driven.
>>
>> * More easily portable to different ARM platforms.  Xen will present a
>> common interface to hide platform idiosyncrasies.
>> --> More widespread embedded use Cog.
>> http://www.slideshare.net/xen_com_mgr/free-rtos-xensummit
>>
>> * Run databases in their own domain.
>>
>> * Isolate plugin memory space. Rather than linking plugins into Cog,
>> potentially threatening its stability, run them in a separate domain and
>> use high-speed inter-domain shared memory.
>> http://www.cs.ucsb.edu/~nanli/projects/cs270.pdf
>>
>> * Have a cloud community of thousands** of Images running each in their
>> own domain, communicating with high-speed inter-domain shared memory.
>>    ** presumptive figure - would need to check scaling ability
>>
>> * Data sharing on high density hosting providers. Not sure if its there
>> yet, but may get Copy-On-Write
>> http://www-archive.xenproject.org/files/xensummit_fall07/18_GregorMilos.pdf
>>
>>
>> * In very large Images, which Spur will allow, perhaps a better way to
>> snapshot an Image is via virtual machine snapshotting using Copy-On.
>> http://www.cercs.gatech.edu/tech-reports/tr2010/git-cercs-10-05.pdf
>>
>> * Rather than convert the Cog/Image to be multi-threaded, could have
>> many grid nodes running their own domain, communicating with high speed
>> shared memory.
>> http://www.cs.ucsb.edu/~nanli/projects/cs270.pdf
>> http://osnet.cs.binghamton.edu/publications/hines07memx.pdf
>>
>> * AMD's hardware virtualisation eliminated a Ring and also segmented
>> memory protection used by virtualisation to separate address spaces, so
>> 64-bit AMD can be slower.  Eliminating the OS kernel layer probably
>> bypasses this limitation since there would be only two levels -
>> Hypervisor & Cog.
>> https://lse.epita.fr/data/2011-lse-summer-week/xen.pdf
>>
>> * Its where all the cool kids hang out --> marketing opportunity
>>
>
> * Security versus lines of code...
>
>    Linux = 15.9 million (3.6 Kernel)
>    OSX = 86 million
>    FreeBSD = 6.25 million
>    OpenSolaris = 10 million
>    Windows = 64 million
>
>    Xen-ARM = 90,000
>    Xen-x86 = 140,000
>
> https://www.google.com/fusiontables/DataSource?docid=12Eg7Ru1CH5MIjomcfjaGKe2qBd9ynmSQni5CAioK
>
>
> http://wiki.xen.org/wiki/Xen_ARM_with_Virtualization_Extensions_whitepaper
>
>
>
>
>
>>
>>
>>>
>>> - what infrastructure is missing?  (e.g. what key drivers need
>>> implementing?)
>>
>> I'll need to understand Cog more. As a start, here are some porting and
>> API docs...
>>
>> http://events.linuxfoundation.org/sites/events/files/slides/Porting%20FreeBSD%20on%20Xen%20on%20ARM%20.pdf
>>
>>
>> http://oss.org.cn/ossdocs/server_storage/xen/interface/interface.html
>>
>> http://www.informit.com/articles/article.aspx?p=1160234&seqNum=4
>>
>> https://lse.epita.fr/data/2011-lse-summer-week/xen.pdf
>>
>> https://github.com/cloudius-systems/osv/blob/master/drivers/xenfront-xenbus.cc
>>
>>
>> cheers -ben
>>
>>>>
>>>> [1] http://pillarhub.pharocloud.com/hub/mikefilonov/pharonos
>>>> [2] http://blog.codinghorror.com/understanding-user-and-kernel-mode/
>>>> [3] http://www.daemonology.net/freebsd-on-ec2/
>>>> [4] http://wiki.netbsd.org/amazon_ec2/amis/
>>>> [5] http://wiki.netbsd.org/amazon_ec2/build_your_own_ami/
>>>> [6] http://wiki.xen.org/wiki/Virtualization_Spectrum
>>>> [7] http://wiki.xen.org/wiki/FrontendDriver
>>>> [8] http://wiki.xen.org/wiki/Cloud_Operating_Systems
>>>
>>>
>>> Eliot (phone)
>>>
>>
>>
>>
>>
>
>
Very interesting thread.


You can also add Docker, an epitech guy's startup
https://xen-orchestra.com/blog/xen-and-docker-for-the-best/


And about the "key economic and managerial benefits" from this kind of 
system : on-demand "elasticity" : to pay only for needed cpu, and to 
avoid the slashsdot effect
link from Sven: http://zerg.erlangonxen.org/
100 ms to boot a VM gest to serve an http request and then exit ... very 
elastic


The PharoNOS project looks great.

I just tried it on an ubuntu+xen and windosw8.1+hyperv dual boot 
desktop, and it runs better on a guest VM with one core and 1GB RAM than
on the desktop host with 12 cores and 16 GB,  on both boot 
configurations (almost every times not always)
funny

test is 0 tinyBenchmarks with a Pharo 3.0 VM:

config 1
Host = Ubuntu 14.04 + Xen 4.4.1, core i7  6 cores, 12 vcpu , 16 GB
Guest VM = PharoNOS (TinyCorelinux, 1 vcpu, 1GB)
(1) Ubuntu 14.04 on host : '1,523,809,523 bytecodes/sec; 219,893,944 
sends/sec'
(2) PharoNOS Xen VM guest : 1,728,270,042 bytecodes/sec; 224,211,429 
sends/sec

config 2: same hardware, booting on Windows8.1 pro with HyperV hypervisor:
Host = Windows 8.1 pro, 12 vcpu, 16 GB
Guest VM = PharoNOS , 1 vcpu, 1GB
(3) Host, Win 8.1 pro:   '1,595,015,576 bytecodes/sec; 206,330,393 
sends/sec'
(4) PharoNOS Hyper-V VM guest :  1,651,612,903 bytecodes/sec; 
201,435,694 sends/sec

Unfortunatly, I can't test it much, I have no network on my Xen vms for 
now (it will come), I tested the available iso of the PharoNOS project



Regards,

Alain