转自： https://specs.openstack.org/openstack/openstack-user-stories/user-stories/proposed/complex-instance-placement.html

This work is licensed under a Creative Commons Attribution 3.0 Unported License.http://creativecommons.org/licenses/by/3.0/legalcode

Complex Instance Placement

Problem description

Problem Definition

An IP Multimedia Subsystem (IMS) core [2] is a key element of Telco infrastructure, handling VoIP device registration and call routing. Specifically, it provides SIP-based call control for voice and video as well as SIP based messaging apps.

An IMS core is mainly a compute application with modest demands on storage and network - it provides the control plane, not the media plane (packets typically travel point-to-point between the clients) so does not require high packet throughput rates and is reasonably resilient to jitter and latency.

As a core Telco service, the IMS core must be deployable as an HA service capable of meeting strict Service Level Agreements (SLA) with users. Here HA refers to the availability of the service for completing new call attempts, not for continuity of existing calls. As a control plane rather than media plane service the user experience of an IMS core failure is typically that audio continues uninterrupted but any actions requiring signalling (e.g. conferencing in a 3rd party) fail. However, it is not unusual for client to send periodic SIP “keep-alive” pings during a call, and if the IMS core is not able to handle them the client may tear down the call.

An IMS core must be highly scalable, and as an NFV function it will be elastically scaled by an NFV orchestrator running on top of OpenStack. The requirements that such an orchestrator places on OpenStack are not addressed in this use case.

Opportunity/Justification

Currently OpenStack supports basic workload affinity/anti-affinity using a concept called server groups. These allow for creation of groups of instances whereby each instance in the group has either affinity or anti-affinity (depending on the group policy) towards all other instances in the group. There is however no concept of having two separate groups of instances where the instances in the group have one policy towards each other, and a different policy towards all instances in the other group.

Additionally there is no concept of expressing affinity rules that can control how concentrated the members of a server group can be - that is, how tightly packed members of a server group can be onto any given hosts. For some applications it may be desirable to pack tightly, to minimise latency between them; for others, it may be undesirable, as then the failure of any given host can take out an unacceptably high percentage of the total application resources. Such requirements can partially be met with so called “soft” affinity and anti-affinity rules (if implemented) but may require more advanced policy knobs to set how much packing or spread is too much.

Although this user story is written from a particular virtual IMS use case, it is generally applicable to many other NFV applications and more broadly to any applications which have some combination of:

Performance requirements that are met by packing related workloads; or
Resiliency requirements that are met by spreading related workloads

Requirements Specification

Use Cases

As a communication service provider, I want to deploy a highly available IMS core as a Virtual Network Function running on OpenStack so that I meet my SLAs.
As an enterprise operator, I want to deploy my traditional database server shards such that they are not on the same physical nodes so that I avoid a service outage due to failure of a single node.

Usage Scenarios Examples

Project Clearwater [3] is an open-source implementation of an IMS core designed to run in the cloud and be massively scalable. It provides P/I/S-CSCF functions together with a BGCF and an HSS cache, and includes a WebRTC gateway providing interworking between WebRTC & SIP clients.

Requirements

The problem statement above leads to the following requirements.

Compute application

OpenStack already provides everything needed; in particular, there are no requirements for an accelerated data plane, nor for core pinning nor NUMA.
HA

Project Clearwater itself implements HA at the application level, consisting of a series of load-balanced N+k pools with no single points of failure [4].

To meet typical SLAs, it is necessary that the failure of any given host cannot take down more than k VMs in each N+k pool. More precisely, given that those pools are dynamically scaled, it is a requirement that at no time is there more than a certain proportion of any pool instantiated on the same host. See Gaps below.

That by itself is insufficient for offering an SLA, though: to be deployable in a single OpenStack cloud (even spread across availability zones or regions), the underlying cloud platform must be at least as reliable as the SLA demands. Those requirements will be addressed in a separate use case.
Elastic scaling

An NFV orchestrator must be able to rapidly launch or terminate new instances in response to applied load and service responsiveness. This is basic OpenStack nova function.
Placement zones

In the IMS architecture there is a separation between access and core networks, with the P-CSCF component (Bono - see [4]) bridging the gap between the two. Although Project Clearwater does not yet support this, it would in future be desirable to support Bono being deployed in a DMZ-like placement zone, separate from the rest of the service in the main MZ.

Gaps

The above requirements currently suffer from these gaps:

Affinity for N+k pools

An N+k pool is a pool of identical, stateless servers, any of which can handle requests for any user. N is the number required purely for capacity; k is the additional number required for redundancy. k is typically greater than 1 to allow for multiple failures. During normal operation N+k servers should be running.

Affinity/anti-affinity can be expressed pair-wise between VMs, which is sufficient for a 1:1 active/passive architecture, but an N+k pool needs something more subtle. Specifying that all members of the pool should live on distinct hosts is clearly wasteful. Instead, availability modelling shows that the overall availability of an N+k pool is determined by the time to detect and spin up new instances, the time between failures, and the proportion of the overall pool that fails simultaneously. The OpenStack scheduler needs to provide some way to control the last of these by limiting the proportion of a group of related VMs that are scheduled on the same host.

External References

Rejected User Stories / Usage Scenarios

None.

Glossary

NFV - Networks Functions Virtualisation, see http://www.etsi.org/technologies-clusters/technologies/nfv
IMS - IP Multimedia Subsystem
SIP - Session Initiation Protocol
P/I/S-CSCF - Proxy/Interrogating/Serving Call Session Control Function
BGCF - Breakout Gateway Control Function
HSS - Home Subscriber Server
WebRTC - Web Real-Time-Collaboration