Hi, On Aug 18, 2005, at 11:51, Pekka Nikander wrote:
"Revisitation" means being able to support larger virtual topologies than you have physical resources for, e.g., deploying a 100-node ring on ten machines in a lab. This is done by hosting multiple virtual nodes on a single physical node. (You "revisit" a physical node multiple times for one overlay - the term may not be perfect.)Many existing systems that deploy virtual networks can only deploy virtual nets that are at most as large as the underlying physical resources, i.e., they can arrange a subset of the physical nodes into arbitrary topologies by creating tunnels, but they cannot create virtual topologies that have more nodes than available physical machines.So far I think I understand.Revisitation is similar to VM, in the sense that both enable the creation of virtual resources of apparently larger capacity than the underlying physical resource. This more completely decouples the two, increasing the power of the virtual abstraction. AFAIK the X-Bone is still the only virtual nets architecture that support this.Hmmm. Are you trying to say that since we have so few real IP addresses available, we need this kind of virtualisation to give more IP addresses per host? If so, I don't understand why? What do you need that many IP addresses for each host?
no, this has nothing to do with IP addresses or IP address space. The motivation was to make the deployable virtual networks as independent from the available physical resources as possible. Without revisitation, virtual networks can only have less or equal the number of nodes than the underlying deployment system. With revisitation, this restriction is lifted.
To use the VM analogy again, a VM without "revisitation" - i.e., paging, would provide protected virtual address spaces to processes, but the virtual address spaces would only ever be as large as the physical memory present. With paging, this restriction is lifted, increasing the usefulness of the VM abstraction immensely.
(Aside: While VM virtualizes address spaces, virtual networks virtualize networks, not the address space the network layer uses. I'm not sure if this is where our misunderstanding comes from; mentioning it just in case.)
(A different aspect of this decoupling between physical and virtual resource would be replication of a single virtual node onto multiple physical machines for fault tolerance. This isn't in X-Bone yet, AFAIK.)I still completely fail to see the utility value of this for anything else but emulation kind of work. Or are you trying to split the two roles of IP addresses by creating a layer of indirection within IP, using virtual addresses as identifiers and physical addresses as locators?
The usefulness lies in removing restrictions on the kinds of virtual networks that are deployable. To deploy a 100-node ring on Planetlab, I first need to check if Planetlab has 100 nodes right now. If it doesn't, then I can't. With the virtual Internet architecture of the X-Bone, I don't need to care how many physical nodes are available to host my virtual network. Similar to how a paging VM supports larger virtual memory spaces than available physical memory.
We've used this capability for fault tolerance purposes, e.g., transparently remapping virtual hosts between physical hosts during the lifetime of a virtual network.
Lars -- Lars Eggert NEC Network Laboratories
smime.p7s
Description: S/MIME cryptographic signature
_______________________________________________ Int-area mailing list [email protected] https://www1.ietf.org/mailman/listinfo/int-area
