openstack team mailing list archive

["Bret Piatt" <bret@xxxxxxxxxxxxx>]

These numbers seem very large today but a decade from now they won't at all.
We don't need to design everything for the requirements in 2020 but we
should have some view on where things are going.  More swag calculations...

In 2020...
60,000,000 servers worldwide (I'm assuming this doesn't grow super quickly
as instead of more, we're making more powerful with additional cores/RAM per
server)
50% of those servers are in "cloud" so 30,000,000
I'll call a "major cloud player" somebody with 10% market share, so
3,000,000 servers
For geographic spread I'll assume 10-15 facilities so 200,000-300,000
servers per facility

Each facility may be broken up into zones so the question is at what point
do we make the cross-over from intra-cloud to inter-cloud?  If it is at the
zone level we're looking at 20-50k servers per zone with increasing VM
density so 50-100 VMs per server for a total of 1 million to 5 million VMs.
At each point in the architecture where the cost model changes for inter-VM
traffic we need to have a logical construct in place.  Potentially
zone->facility->region->world is enough granularity, we may need
zone->facility->metro area->region->world.  Where we switch from intra-cloud
(singular management domain) to inter-cloud (connected through a cloud
directory service with federation) needs to take into account the overhead
for brokering the additional complexity of inter-cloud communication versus
the overhead of a larger singular management domain.

Bret Piatt
OpenStack.org
Twitter: @bpiatt / Mobile 210-867-9338
Open Source Cloud Infrastructure: http://www.openstack.org



-----Original Message-----
From: openstack-bounces+bret=openstack.org@xxxxxxxxxxxxxxxxxxx
[mailto:openstack-bounces+bret=openstack.org@xxxxxxxxxxxxxxxxxxx] On Behalf
Of Erik Carlin
Sent: Thursday, December 30, 2010 12:15 PM
To: Rick Clark
Cc: openstack@xxxxxxxxxxxxxxxxxxx
Subject: Re: [Openstack] Some insight into the number of instances Nova
needs to spin up...

I suggest we consider the limits of a single nova deployment, not across all
regions.  To Pete's point, at a certain scale, people will break into
parallel, independent nova deployments.  A single, global, deployment
becomes untenable at scale.

I agree with Pete that 1M hosts (and 45M Vms) is a bit out of whack for a
single nova deployment.  As a frame of reference, here are a couple of links
that estimate total server count by the big boys:

http://www.datacenterknowledge.com/archives/2009/05/14/whos-got-the-most-we
b-servers/
http://www.intac.net/a-comparison-of-dedicated-servers-by-company_2010-04-1
3/

Google is the largest and they are estimated to run ~1M+ servers.
Microsoft is ~500K+.  Facebook is around 60K.  This link
(http://royal.pingdom.com/2009/08/24/google-may-own-more-than-2-of-all-serv
ers-in-the-world/) is a few years old and puts the total worldwide server
count at 44M.

I submit that setting the nova limit to match Google's total server count
and 1/44th of the total worldwide server count is overkill.

The limits I suggest below are not per AZ, but per nova deployment (there
could be multiple AZs inside of a deployment).  I think we may need to
clarify nomenclature (although it may just be me since I haven't been too
engaged in these discussions to date).  I know at the last design summit it
was decided to call everything a "zone".

Erik
 

On 12/30/10 11:42 AM, "Rick Clark" <rick@xxxxxxxxxxxxx> wrote:

>Actually it was 1M hosts and  Ivthink 45 million vms.  It was meant to 
>be across all regions.  Jason Seats set the number arbitrarily, but it 
>is a good target to not let us forget about scaling while we design.
>
>I think eventually all loads will be more ephemeral.  So, I think I 
>agree with your numbers, if you are talking about a single availability 
>zone.
>
>On 12/30/2010 11:25 AM, Erik Carlin wrote:
>> You are right.  The 1M number was VMs not hosts.  At least, that was 
>>from  one scale discussion we had within Rackspace.  I'm not sure what 
>>the  "official" nova target limits are and I can't find anything on 
>>launchpad  that defines it.  If there is something, could someone 
>>please send me a  link.
>> 
>> I'm am certain that Google can manage more than 10K physical servers 
>> per DC. Rackspace does this today.
>> 
>> If I combine what I know about EC2 and Cloud Servers, I would set the 
>>ROM  scale targets as:
>> 
>> ABSOLUTE
>> 1M VMs
>> 50K hosts
>> 
>> RATE
>> 500K transactions/day (create/delete server, list servers, resize 
>>server,  etc. - granted, some are more expensive than others)  That 
>>works out to ~21K/hr but it won't be evenly distributed.  To allow  
>>for peak, I would say something like 75K/hour or ~21/sec.
>> 
>> Thoughts?
>> 
>> Erik
>> 
>> 
>> On 12/30/10 9:20 AM, "Pete Zaitcev" <zaitcev@xxxxxxxxxx> wrote:
>> 
>>> On Wed, 29 Dec 2010 19:27:09 +0000
>>> Erik Carlin <erik.carlin@xxxxxxxxxxxxx> wrote:
>>>
>>>> The 1M host limit still seems reasonable to me. []
>>>
>>> In my opinion, such numbers are completely out of whack. Google's 
>>>Chubby  article says that the busiest Chubby has 90,000 clients (not 
>>>hosts!) and  the biggest datacenter has 10,000 systems. They found 
>>>such numbers  pushing the border of unmanageable. Granted they did 
>>>not use  virtualization, but we're talking the number of boxes in 
>>>both cases.
>>>
>>> So to reach 1M hosts in a Nova instance you have to have it manage
>>> 100 datacenters. There are going to be calls for federation of Novas 
>>> long before this number is reached.
>>>
>>> Sustaining a high flap rate is a worthy goal and will have an 
>>> important practical impact. And having realistic sizing ideas is 
>>> going to help it.
>>>
>>> -- Pete
>> 
>> 
>> 
>> Confidentiality Notice: This e-mail message (including any attached 
>>or  embedded documents) is intended for the exclusive and confidential 
>>use of the  individual or entity to which this message is addressed, 
>>and unless otherwise  expressly indicated, is confidential and 
>>privileged information of Rackspace.
>> Any dissemination, distribution or copying of the enclosed material 
>>is prohibited.
>> If you receive this transmission in error, please notify us 
>>immediately by e-mail  at abuse@xxxxxxxxxxxxx, and delete the original 
>>message.
>> Your cooperation is appreciated.
>> 
>> 
>> _______________________________________________
>> Mailing list: https://launchpad.net/~openstack
>> Post to     : openstack@xxxxxxxxxxxxxxxxxxx
>> Unsubscribe : https://launchpad.net/~openstack
>> More help   : https://help.launchpad.net/ListHelp
>
>



Confidentiality Notice: This e-mail message (including any attached or
embedded documents) is intended for the exclusive and confidential use of
the individual or entity to which this message is addressed, and unless
otherwise expressly indicated, is confidential and privileged information of
Rackspace. 
Any dissemination, distribution or copying of the enclosed material is
prohibited.
If you receive this transmission in error, please notify us immediately by
e-mail at abuse@xxxxxxxxxxxxx, and delete the original message. 
Your cooperation is appreciated.


_______________________________________________
Mailing list: https://launchpad.net/~openstack
Post to     : openstack@xxxxxxxxxxxxxxxxxxx
Unsubscribe : https://launchpad.net/~openstack
More help   : https://help.launchpad.net/ListHelp