Re: [alto] How Data Center Virtualization influence ALTO mechanism.

[stefano previdi <[e-mail]>]

On Oct 9, 2010, at 9:18 AM, Y.J. GU wrote:

> Hi all,
> I was thinking about how Data Center Virtualization and Virtual  
> Machine(VM) Migration will influence ALTO mechanism.
>
> Current ALTO Protocol defines clustering of peers according to their  
> IP Addresses. E.g. peers in same subnet will be classified into same  
> PID, and path cost will indicate the cost within and between PIDs,  
> which is also actually based on IP Addresses.

the methods of grouping are orthogonal to ALTO protocol specification.

There are ALTO implementations that allow address/prefixes grouping
relaxed from pure IP aggregation. The fact that you use IP addresses
in the protocol doesn't mean that locality is solely based on
address/mask pairs.

It is mostly a policy definition by the ALTO and infrastructure
operator.


> In the current world, peers are partitioned by IP subnet. While  
> considering virtual machines migration, there might be more  
> interesting things to think of.
>
> In Data Center operation, one basic consensus is 'When Virtual  
> Machines move from one site to another, the IP Addresses will not  
> change, so that the existing service connection will not be  
> broken'.  VMs can migrate to arbitrary site, not under the control  
> and knowledge of ISP. For example, some VMs in Data Center A(IP  
> subnet 198.1.1.0) move to Data Center B (IP subnet 210.1.1.0). IP- 
> based, Vms are closer to DC-A. Physically, these VMs are much closer  
> to hosts in DC-B. However things are not so easy, especially  
> considering how these VMs are routed. Current ALTO may give wrong  
> cost ranking.


that is true. ALTO relies on accurate infrastructure/topology
information. It can be derived from lower layers (routing and below)
or inferred by policy DBs.


> VMs may migrate under, but not limited to, these situations: 1) to  
> save electricity power, 2) disaster recovery, 3) customer prefer  
> another Data Center, 4) company extension, etc. In the end, the  
> internet will not be a regular world partitioned by IP Addresses.


the "swamp" already validated the theory...


> Does anyone think this is an interesting aspect to study?

probably yes but I'm not sure I see the protocol implication.

s.