| Network Working Group | K. Kompella |
| Internet-Draft | Y. Rekhter |
| Intended status: Standards Track | Juniper Networks |
| Expires: April 23, 2013 | T. Morin |
| France Telecom - Orange Labs | |
| D.L. Black | |
| EMC Corporation | |
| October 22, 2012 |
Signaling Virtual Machine Activity to the Network Virtualization Edge
draft-kompella-nvo3-server2nve-01
This document proposes a simplified approach for provisioning the networking parameters related to Virtual Machine creation, migration and termination on servers. The idea is to provision the server, then have the server signal the requisite parameters to the relevant network device(s). Such an approach reduces the workload on the provisioning system and simplifies the data model that the provisioning system needs to maintain. Furthermore, it is more resilient to topology changes in server-network connectivity, for example, reconnecting a server to a different network port or switch.
This Internet-Draft is submitted in full conformance with the provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering Task Force (IETF). Note that other groups may also distribute working documents as Internet-Drafts. The list of current Internet- Drafts is at http:/⁠/⁠datatracker.ietf.org/⁠drafts/⁠current/⁠.
Internet-Drafts are draft documents valid for a maximum of six months and may be updated, replaced, or obsoleted by other documents at any time. It is inappropriate to use Internet-Drafts as reference material or to cite them other than as "work in progress."
This Internet-Draft will expire on April 23, 2013.
Copyright (c) 2012 IETF Trust and the persons identified as the document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http:/⁠/⁠trustee.ietf.org/⁠license-⁠info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Code Components extracted from this document must include Simplified BSD License text as described in Section 4.e of the Trust Legal Provisions and are provided without warranty as described in the Simplified BSD License.
To create a Virtual Machine (VM) on a server in a data center, one must specify parameters for the CPU, storage, network and appliance aspects of the VM. At a minimum, this requires provisioning the server that will host the VM, and the Network Virtualization Edge (NVE) that will implement the virtual network for the VM. Similar considerations apply to live migration and terminating VMs. This document proposes mechanisms whereby a server can be provisioned with all of the paramters for the VM, and the server in turn signals the networking aspects to the NVE. The NVE may be located on the server or in an external network switch that may be directly connected to the server or accessed via an L2 (Ethernet) LAN or VLAN. The following subsections capture the abstract sequence of steps for VM creation, live migration and deletion.
This subsection describes an abstract sequence of steps involved in creating a VM and make it operational. The following steps are intended as an illustrative example, not as prescriptive text; the goal is to capture sufficient detail to set a context for the signaling described in Section 5.
Creating a VM requires:
While shown as a numbered sequence above, some of these steps may be concurrent (e.g., server, storage and network provisioning for the new VM may be done concurrently).
Steps 1 and 2 are primarily information gathering. For Steps 3 to 8, the provisioning system talks actively to servers, network switches, storage and appliances, and must know the details of the physical server, network, storage and appliance connectivity topologies. Step 4 is typically done using just provisioning, whereas Steps 5 and 6 may be a combination of provisioning and other techniques. Steps 4 to 6 accomplish the task of provisioning the network for a VM, the result of which is a Data Center Virtual Private Network (DCVPN) overlaid on the physical network.
This document focuses on the case where the network elements in Step 4 are not co-resident with the server, and shows how the provisioning in Step 4 can be replaced by signaling between server and network, using information from Step 3. This document also shows how Step 4 can interact seamlessly with some of the realizations of Steps 5 and 6.
This subsection describes an abstract sequence of steps involved in live migration of a VM. Live migration is sometimes referred to as "hot" migration, in that from an external viewpoint, the VM appears to continue to run while being migrated to another server (e.g., TCP connections generally survive this class of migration). In contrast, suspend/resume (or "cold") migration consistes of suspending VM execution on one server and resuming it on another. The following live migration steps are intended as an illustrative example, not as prescriptive text; the goal is to capture sufficient detail to set a context for the signaling described in Section 5.
For simplicity, this set of abstract steps assumes shared storage, so that the VM's storage is accessible to the source and destination servers. Live migration of a VM requires:
While shown as a numbered sequence above, some of these steps may be concurrent (e.g., moving the VM and associated network changes).
Step 1 is primarily information gathering. For Steps 2, 3, 10 and 11, the provisioning system talks actively to servers, network switches and appliances, and must know the details of the physical server, network and appliance connectivity topologies. Steps 4 and 5 are usually handled directly by the servers involved. Steps 6 to 9 may be handled by the servers (e.g., a gratuitous ARP or RARP from the destination server may accomplish all four steps) or other techniques.
This document focuses on the case where the network elements are not co-resident with the server, and shows how the provisioning in Step 3 and the deprovisioning in Step 10 can be replaced by signaling between server and network, using information from Step 3. This document also shows how Step 4 can interact seamlessly with some of the realizations of Steps 5 and 6.
This subsection describes an abstract sequence of steps involved in termination of a VM, also referred to as "powering off" a VM. The following termination steps are intended as an illustrative example, not as prescriptive text; the goal is to capture sufficient detail to set a context for the signaling described in Section 5.
Termination of a VM requires:
While shown as a numbered sequence above, some of these steps may be concurrent (e.g., network deprovisioning and VM deletion).
Steps 1, 2 and 4 are handled by the server, based on instructions from the provisioning system. For Step 3, the provisioning system talks actively to servers, network switches, storage and appliances, and must know the details of the physical server, network, storage and appliance connectivity topologies.
This document focuses on the case where the network elements in Step 3 are not co-resident with the server, and shows how the deprovisioning in Step 3 can be replaced by signaling between server and network.
The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in [RFC2119].
The following acronyms are used:
The goal of provisioning networks for VMs is to create an "isolation domain" wherein a group of VMs can talk freely to each other, but communication to and from VMs outside that group is restricted (either prohibited, or mediated via a router, a firewall or other network gateway). Such an isolation domain, sometimes called a Closed User Group, here will be called a Data Center Virtual Private Network (DCVPN). The network elements on the outer border or edge of the overlay portion of a Virtual Network are called Network Virtualization Edges (NVEs).
A DCVPN is assigned a global "name" that identifies it in the management plane; this name is unique in the scope of the data center, but may be unique across several cooperating data centers. A DCVPN is also assigned an identifier unique in the scope of the data center, the Virtual Network Group ID (VNID). The VNID is a control plane entity. A data plane tag is also needed to distinguish different DCVPNs' traffic; more on this later.
For a given VM, the NVE can be classified into two parts: the network elements to which the VM's server is directly connected (the local NVE or l-NVE), and those to which peer VMs are connected (the remote NVE or r-NVE). In some cases, the l-NVE is co-resident with the server hosting the VM; in other cases, the l-NVE is separate (distributed l-NVE). The latter case is the one of interest in this document.
A created VM is added to a DCVPN through Steps 4 to 6 in section Section 1.1 which can be recast as follows. In Step 4, the l-NVE(s) are informed about the VM's VNID, network addresses and policies, and the lNVE and server agree on how to distinguish traffic for different DCVPNs from and to the server. In Step 5 the relevant r-NVE elements and the addresses of their VMs are discovered. In Step 6, the r-NVE(s) are informed of the presence of the new VM and obtain its addresses.
Once a DCVPN is created, the next steps for network provisioning are to create and apply policies such as for QoS or access control. These occur in three flavors: policies for all VMs in the group, policies for individual VMs, and policies for communication across DCVPN boundaries.
DCVPNs are often realized as Ethernet VLAN segments. A VLAN segment satisfies the communication properties of a DCVPN. A VLAN also has data plane mechanisms for discovering network elements (Layer 2 switches, aka bridges) and VM addresses. When a DCVPN is realized as a VLAN, Step 4 requires provisioning both the server and l-NVE with the VLAN tag that identifies the DCVPN. Step 6 requires provisioning all involved network elements with the same VLAN tag. Address learning is done by flooding, and the announcement of a new VM is typically by a "gratuitous ARP".
While VLANs are familiar and well-understood, they fail to scale on several dimensions. Underlying VLANs is a Layer 2 infrastructure. The number of independent VLANs in a Layer 2 domain is limited by the size of the VLAN tag. Data plane techniques (flooding and broadcast) are another source of serious concern as the overall size of the network grows.
There are several scalable realizations of DCVPNs that address the isolation requirements of DCVPNs as well as the need for a scalable substrate for DCVPNs and the need for scalable mechanisms for NVE and VM address discovery. While these are not the goal of this document, a secondary goal of this document is to show how the signaling that replaces Step 4 can seamlessly interact with several of these realizations of DCVPNs.
VLAN tags (VIDs) will be used as the data plane tag to distinguish traffic for different DCVPNs' between a server and its l-NVE. Note that, as used here, VIDs only have local significance between server and NVE, not to be confused with the notion of VLANs, which are a data center-wide concept. Data plane tags between l-NVE and r-NVE depends on the encapsulation mechanism among the NVE; the l-NVE is expected to map between VIDs and intra-NVE tags in both directions.
For VM creation as described in section Section 1.1, Step 3 provisions the server; Steps 4 and 5 provision the l-NVE elements; Step 6 provisions the r-NVE elements.
In some cases, the l-NVE elements live within the server; in this case, Steps 3 and 4 are "single-touch" in that the provisioning system only needs to talk to the server, and both CPU and network parameters can be applied by the server. However, in other cases, the l-NVE is separate from the server, requiring that the provisioning system talk independently to both the server and lNVE. This scenario, which we call "distributed local NVE", is the one considered in this document. This document resurrects "single-touch" provisioning in the distributed lNVE case.
The approach here is to provision the server, then have the server signal the requisite parameters to the l-NVE. Such an approach reduces the workload on the provisioning system, allowing it to scale both in the number of elements it can manage, as well as the rate at which it can process changes. It also simplifies the data model that the provisioning system needs to have; in particular, the provisioning system does not have to maintain a full, up-to-date map of server to network connectivity. Furthermore, it is more resilient to topology changes in server-network connectivity that have not yet been transmitted to the provisioning system. For example, if a server is reconnected to a different port or a different l-NVE to recover from a malfunctioning port, the server can contact the new l-NVE over the new port without the provisioning system being aware of the change.
While the current document focuses on provisioning networking parameters via signaling, future extensions may address the provisioning of storage and middle-box parameters in a similar fashion. Companion documents will describe how NVEs to which peer VMs are connected can get the required networking information via signaling rather than by provisioning and/or other means.
There are three common operations in a virtualized data center: creating a VM; migrating a VM from one physical server to another; and terminating a VM. Creating a VM requires "associating" it with its DCVPN and "activating" that association; decommissioning a VM requires "deactivating" the VM's association with the DCVPN and then "dissociating" the VM from its DCVPN. Moving a VM consists of associating it with its DCVPN in its new location, then dissociating it from its old location. . The deactivation operation is often implicit in another operation, but is called out here for symmetry and completeness.
For each VM association operation, a subset of the following information is needed from server to l-NVE:
Activate and deactivate are dataplane operations that reference the VID, and additionally provide authentication, table type and address entries information. When an activate is realized via a "gratuitous ARP" in the data plane, the VID is in the Ethernet header, and all of the other parameters are obtained by mapping the VID and the port on which the frame containing it was received to information established by a prior associate operation.
Realizations of DCVPNs include, among others, E-VPNs ([I-D.ietf-l2vpn-evpn]), IP VPNs ([RFC4364]), NVGRE ([I-D.sridharan-virtualization-nvgre], TRILL ([RFC6325]), VPLS ([RFC4761], [RFC4762]), and VXLAN ([I-D.mahalingam-dutt-dcops-vxlan]). The table type implicitly defines whether forwarding at the NVE for the DCVPN is at Layer 2 or Layer 3 or both.
Typically, for the pre-associate and associate messages, all the information except hold time would be needed. For the dissociate message, all the above information except VID and table type would be needed.
Operations are stateful, that is, they remain in place until superceded by another operation. For example, on receiving an associate message, an NVE is expected to create and maintain the DCVPN table for a VM until the NVE receives a dissociate message to remove the table. A separate liveness protocol may be run between server and NVE to let each side know that the other is still operational; if the liveness protocol fails, each side may remove all state installed in response to messages from the other.
In the descriptions below, we assume that the NVE layer provides a mechanism for control plane distribution of VM addresses, as opposed to doing this in the data plane. If this is not the case, NVE elements can skip the parts of the procedures below that involve address distribution.
As VIDs are local to server-NVE communication, in fact to a specific port connecting these two elements, a mapping table containg 4-tuples of the following form will prove useful to the NVE:
<VID, port, VNID, VM address entries>
The procedures below assume that the NVE systematically reorders the provided VM address entries before inserting or looking up entries in this mamping table.
Note that valid values of VID are from 1 to 4094, inclusive. A value of 0 is used to mean "unassigned". When a VID can be shared by more than one VM, it is necessary to reference-count entries in this table. Entries in this table have multiple uses:
When a VM is instantiated on a server, it is assigned a VNID, VM addresses and a table type for the DCVPN. The VM addresses may be any of IPv4, IPv6 and MAC addresses. There may also be network policies specific to the VM. To connect the VM to its DCVPN, the server signals these parameters to the l-NVE via an "associate" operation followed by an "activate" operation to put the parameters into use. (Note that the l-NVE may consist of more than one device.)
On receiving an associate message on port P from server S, an NVE device does the following:
After a successful associate, the network has been provisioned (at least in the local NVE) for the VM's traffic, but forwarding has not been enabled. On receiving an activate message on port P from server S, an NVE device does the following (activate is a one-way message that does not have a response):
On receiving a request from the provisioning system to terminate a VM, the server sends a dissociate message to the l-NVE with the hold time set to zero. The dissociate message contains the operation, authentication, VNID, table type, and VM addresses. On receiving the dissociate message on port P from server S, each NVE device L does the following:
NOTE: This sub section has not been updated from the -00 version of this draft; it will be updated in the forthcoming -02 version. The set of VM migration steps are known to be incomplete, material on concurrent actions and race conditions (based on list discussion) should be added and new step PA.5 is anticipated to need generalization to encompass control planes that may not push all addressing changes to all relevant rNVEs. Please ignore the text in this subsection and beyond - this document is a draft and the authors are working on it.
Let's say that a VM is to be migrated from server S (connected to lNVE device L) to server S' (connected to lNVE device L'). The sequence of steps for migration is:
At Step M.1, S' initiates the move, and also sends a pre-associate message to L', including the pre-associate information. The processing of a pre-associate message (PA.1 to PA.7) for L' is the same as that of an associate message (A.1 to A.6), with the following change to step 5.
(*) See Section 6 for some mechanisms for doing this. This is necessary so that L' does not attract traffic to the VM's new location before the migration is complete, yet L knows ahead of time how to send traffic to L' (Step D.2), minimizing traffic loss to the VM when migration is complete.
At step M.2, S initiates the VM copy. If at any time L hears advertisements from L' about how to communicate with the VM in its new location (as unpreferred destinations), L stores that information for use in step D.2.
At step M.3, S terminates the running of the VM on itself, and sends a dissociate message to L with a non-zero hold time (either what the provisioning system sends, or a default value). L processes the dissociate message as above.
There are several options for protocols to use to signal the above messages. One could invent a new protocol for this purpose. One could reuse existing protocols, among them LLDP, XMPP, HTTP REST, and VDP [VDP], a new protocol standardized for the purposes of signaling a VM's network parameters from server to lNVE. Several factors influence the choice of protocol(s); at this time, the focus is on what needs to be signaled, leaving for later the choice of how the information is signaled, and specific encodings.
Procedures to handle failures of the server or of the NVE will be covered in a further revision.
The control plane for a DCVPN manages the creation/deletion, membership and span of the DCVPN ([I-D.narten-nvo3-overlay-problem-statement], [I-D.kreeger-nvo3-overlay-cp]). Such a control plane needs to work with the server-to-nve signaling in a coordinated manner, to ensure that address changes at a local NVE are reflected appropriately in remote NVEs. The details of such coordination will be specified in a companion document.
Many thanks to Amit Shukla for his help with the details of EVB and his insight into data center issues.
| [RFC2119] | Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, March 1997. |
| [VDP] | , , "Edge Virtual Bridging (802.1Qbg) (work in progress)", 2012. |