]> ETRI

218 Gajeong-ro Yuseung-Gu Daejeon 305-700 Korea +82 42 860 1483 seungiklee@etri.re.kr

Korea University

145 Anam-ro, Seongbuk-gu Seoul 136-701 Korea +82 2 3290 4825 shpack@korea.ac.kr

ETRI

218 Gajeong-ro Yuseung-Gu Daejeon 305-700 Korea +82 42 860 4847 mkshin@etri.re.kr

KT

17 Woomyeon-dong, Seocho-gu Seoul 137-792 Korea +82 2 526 5233 eun.paik@kt.com

Intel

Dromore House, East Park Shannon Co. Clare Ireland +353 61 477400 rory.browne@intel.com

General Internet Research Task Force (IRTF) Internet Draft This document specifies problem definition and use cases of NFV resource management in service chaining for path optimization, traffic optimization, failover, load balancing, etc. It further describes design considerations and relevant framework for the resource management capability that dynamically creates and updates network forwarding paths (NFPs) considering resource constraints of NFV infrastructure.

Network Functions Virtualisation (NFV) offers a new way to design, deploy and manage network services. The network service can be composed of one or more network functions and NFV relocates the network functions from dedicated hardware appliances to generic servers, so they can run in software. Using these virtualized network functions (VNFs), one or more VNF forwarding graphs (VNF-FGs; a.k.a. service chains) can be associated to the network service, each of which describes a network connectivity topology, by referencing VNFs and Virtual Links that connect them. One or more network forwarding paths (NFPs) can be built on top of such a topology, each defining an ordered sequence of VNFs and Virtual Links to be traversed by traffic flows matching certain criteria. The network service is instantiated by allocating NFVI resources for VNFs and VLs which constitute the VNF-FGs. Thus, the capacity and performance of the network service depends on the state and attributes of the network resources used for its VNF and VL instances. While this brings a similar problem to the VM placement optimization in a cloud computing environment, it differs as one or more VNF instances are interconnected for a single network service. For example, if one of the VNF instances in the VNF-FG gets failed or overloaded, the whole network service also gets affected. Thus, the VNF instances need to be carefully placed during NS instantiation considering their connectivity within NFPs. They also need to be monitored and dynamically migrated or scaled at run-time to adapt to changes in the resources. The resource management problem in VNF-FGs matters not only to the performance and capacity of network services but also to the optimized use of NFVI resources. For example, if processing and bandwidth burden converges on the VNF instances placed in a specific NFVI-PoP, it may result in scalability problem of the NFV infrastructure. Thus care is encouraged to be taken in distributing load across local and external VNF instances at run-time. This document addresses resource management problem in service chaining to optimize the NS performance and NFVI resource usage. It provides the relevant use cases of the resource management such as traffic optimization, failover, load balancing and further describes design considerations and relevant framework for the resource management capability that dynamically creates and updates NFP instances considering NFVI resource states for VNF instances and VL instances. Note that this document mainly focuses on the resource management capability based on the ETSI NFV framework but also studies contribution points to the work for control plane of SFC architecture .

This document uses the following terms and most of them were reproduced from . Network Functions (NF): A functional building block within a network infrastructure, which has well-defined external interfaces and a well-defined functional behavior. Network service: A composition of network functions and defined by its functional and behavioural specification. NFV Framework: The totality of all entities, reference points, information models and other constructs defined by the specifications published by the ETSI ISG NFV. Virtualised Network Function (VNF): An implementation of an NF that can be deployed on a Network Function Virtualisation Infrastructure (NFVI). NFV Infrastructure (NFVI): The NFV-Infrastructure is the totality of all hardware and software components which build up the environment in which VNFs are deployed. NFVI-PoP: A location or point of presence that hosts NFV infrastructure VNF Forwarding Graph (VNF-FG): A NF forwarding graph where at least one node is a VNF. Network Forwarding Path (NFP): The sequence of hardware/software switching ports and operations in the NFV network infrastructure as configured by management and orchestration that implements a logical VNF forwarding graph "link" connecting VNF "node" logical interfaces. Virtual Link: A set of connection points along with the connectivity relationship between them and any associated target performance metrics (e.g. bandwidth, latency, QoS). The Virtual Link can interconnect two or more entities (e.g., VNF components, VNFs, or PNFs).

This section addresses several issues for considerations in NFV resource management of service chain.

In the NFV framework, network services are realized with NS instantiation procedures at which virtualized NFVI resources are assigned to the VNFs and VLs which constitute VNF-FGs of the network service. The NFVI resources are placed and located along the VNF-FG by NFV Orchestrator (NFVO) dynamically according to: Resource availability, Deployment templates which define resource requirements of NS instances and VNF instances to support KPIs (e.g., capacity and performance) of the network service, and Resource policies which define how to govern NFVI resources for NS instances and VNF instances (e.g., affinity/anti-affinity rules, scaling, and fault management) to support an efficient use of NFVI resources as well as KPIs of the network service. In order to satisfy the deployment templates and resource policies, VNF-FGs of the network services need to be built by considering the state of NFVI resources for VNF instances (e.g., availability, throughput, load, disk usage) and VL instances (e.g., bandwidth, delay, delay variation, packet loss). However, since the NFVI resources are shared by different network services and their deployment constraints are very different from each other, it is required to carefully schedule the NFVI resources for multiple network services to optimize their KPIs.

In NFV, a network service is composed of one or more virtualized network functions which are connected via virtual links along NFPs specified for a traffic flow for the network service. Thus, the performance of a network service is determined by the performance and capability of a coupling of VNF instances and VL instances. For example, if one of the VNF instances or VL instances of an NFP gets failed or overloaded, the whole network service also gets affected. Thus, the VNF instances need to be carefully placed during NS instantiation considering their connectivity within NFPs. This performance coupling can be handled by considering deployment rules for affinity/anti-affinity, geography, or topological locations of VNFs; and QoS of virtual links. Another important factor for virtual links is the inter-connectivity between different NFVI-PoPs, which is an enabler of resource sharing among different NFVI-PoPs. When the VNF instances of a network service are allocated at different NFVI-PoPs, the NFVI-PoP interconnect may be a bottleneck point which needs to be monitored to support KPIs of the service chain.

The NFVI resources for a network service should be allocated and managed according to a NS policy given in the network service descriptor (NSD), which describes how to govern NFVI resources for VNF instances and VL instances to support KPIs of the network service. The examples of NS policy are affinity/anti-affinity, scaling, fault and performance, geography, regulatory rules, NS topology, etc. Since network services may have different NS policies for their own deployment and performance, this may cause resource management difficult within the shared NFVI resources. For network-wide (or NS-wide) resource management, NFVI policy (or network policy) can be also provided. It may describe the resource management policy for optimized use of infrastructure resources rather than the performance of a single network service. The examples of NFVI policy are NFVI resource access control, reservation and/or allocation policies, placement optimization based on affinity and/or anti-affinity rules, geography and/or regulatory rules, resource usage, etc. Multiple administrative domains or subsystems may have different NFVI policies so that it may bring some conflicts when enforcing them in a global infrastructure. There could be a similar problem among NS policies and NFVI policies. Note that the similar topics are being studied in

The performance and capability of NFVI resources may vary in time due to different uses and management policies of the resources. If some changes in the resources make the service quality unacceptable, the VNF instances can be scaled according to the given auto-scaling policies. But it's only for local quality of the VNF. In order to provide optimized KPIs to network services, the NFP instances need to dynamically adapt to the changes of the resource state at run-time. The performance of the whole NFP instance should be measured by monitoring the resource state of VNF instances and VL instances. Based on the monitoring results, some VNF instances may be determined and relocated at different virtualized resources with better performance and capabilities.

In this section, several (but not exhausted) use cases for resource management in service chaining are provided: fail-over, load balancing, path optimization, traffic optimization, and energy efficiency.

For service continuity, failure of a VNF instance needs to be detected and the failed one needs to be replaced with the other one which is available to use as per redundancy policy. Figure 1 presents an example of the fail-over use case. A network service is defined as a chain of VNF-A and VNF-B; and the service chain is instantiated with VNF-A1 and VNF-B1 which are instances of VNF-A and VNF-B, respectively. In the meantime, failure of VNF-B1 is detected so that VNF-B2 replaces the failed one for fail-over of the NFP.

+--------+ +--------+ | VNF-B2 | #| VNF-B2 |### +--------+ +--------+ +--------+ # +--------+ ###| VNF-A1 | _|_ ###| VNF-A1 |# _|_ +--------+ (___) +--------+ (___) ___/ # / \ \ ___/ / \ (___)+---#------+ + ===} (___)+----------+ + # \ ___ / / \ ___ / # (___) (___) # | | # +--------+ +--------+ ######| VNF-B1 |### (failure)--> | VNF-B1 | +--------+ +--------+ ### NFP

The above is in the case where there is a 1+1 or 1:N redundancy scheme. In event that VNF instance overloads before NFVO has time to scale out, or when resources do not permit a scale-out then we can route the service chain deterministically to a remote VNF instance. This adaptation may be revertive or non-revertive dependent on service provider policy and resource availability.

A single VNF instance may be a bottleneck point of a service chain due to its overload. It may affect the performance of the whole service chain consequently so that an NFP instance needs to be built to avoid bottleneck points or maintained to distribute workloads of overloaded VNF instances. With NFVI-PoP Interconnect, service chains can be balanced between NFVI-PoPs in a way that best utilises NFV infrastructure and ensures service integrity. The wide area conditions can be monitored in real-time to provide KPIs, such as BW, delay, delay variation and packet loss per QoS class to the service chaining application which may enable use of external VNF instances when there is an overload or failure condition in the local NFVI-PoP. In this way the service chaining application can make a service chain reroute decision (in the event of failure and/or overload) that is network and platform-aware. The service chaining application understands the state of external VNFs and WAN conditions per QoS class between the local NFVI-PoP and remote NFVI-PoP in real-time.

Traffic for a network service traverses all of the VNF instances and the connecting VL instances given by a NFP instance to reach a target end point. Thus, quality of the network service depends on the resource constraints (e.g., processing power, bandwidth, topological locations, latency) of VNF instances, VL instances including NFVI-PoP interconnects. In order to optimize the path of the network service, the resource constraints of VNF instances and VLs need to be considered at constructing NFPs. Since the resource state may vary in time during the service, NFP instances also need to adapt to the changes of resource constraints of the VNF instances and VL instances by monitoring and replacing them at run-time.

A network operator may provide multiple network services with different VNF-FGs and different flows of traffic traverse between source and destination end-points along the VNF-FGs. For efficiency of resource usage, the NFP instances need to be built by default to localize the traffic flows and to avoid processing and network bottlenecks. It is only in the case of local failure or overload (whereby the NFVO is unable or has not completed a scale-out of on-site resources) that NFP instances would be constructed between NFVI-PoPs. In this case, multiple VNF instances of different NFVI-PoPs need to be considered together at constructing a new NFP instance or adapting one.

Energy efficiency in the network is getting important to reduce impact on the environment so that energy consumption of VNF instances using NFVI resources (e.g., compute, storage, I/O) needs to be considered at NFP instantiation or adaptation. For example, a NFP can be instantiated as to make traffic flows aggregated into a limited number of VNF instances as much as its performance is preserved in a certain level. Policy may vary between centralized or distributed NFV applications, and could include policies for even energy distribution between sites, time-of-day etc.

To derive specific algorithms for use cases discussed in Section 4, an evaluation model for a service chain (or a NFP) needs to be developed, which can address two problems for a given network topology and input parameters (e.g., VL/VNF capacity, incoming traffic flows, etc.) : 1) how much traffic flows pass on each VL instance and 2) how much processing capacity is needed for the installed VNF instance. This section first describes the system model and then presents main objectives for the evaluation model.

The system model considers the following network topology. The network topology under consideration is composed of start/end points and multiple NFVI-PoPs where multiple VNF instances locate. On the other hand, VL instances inter-connect VNF instances in NFVI-PoPs. Start and end points are incoming and outgoing points of traffic flow for a given network service, respectively. Specifically, the amount of incoming traffic flows for a network service (i.e., a VNF-FG) at the start point is given as an input parameter in the model. Under the network topology, the network traffic is processed by one or more VNF instances and delivered via VL instances. Thus, the VNF processing capacity can be defined as the maximum amount of traffic flows that a VNF instance can process according to the resource allocation policies defined in its deployment template. The VL capacity can be also defined as the maximum amount of traffic flows that can pass on a VL instance according to the resource allocation policies defined in the deployment template. In NFV, traffic flows for a VNF-FG should be processed according to the VNF order described in the given VNF-FG. Accordingly, traffic flows at the start point should not be processed by any VNF. Meanwhile, traffic flows at the end point should be processed by all VNFs specified in the given VNF-FG. In a given VNF-FG, VNFs should be individually placed on multiple NFVI-PoPs. Therefore, a decision variable, VNF placement indicator function (VPIF), is defined as: VNF placement indicator function (VPIF): indicator function (i.e., 0 or 1) to represent the location (i.e., a NFVI-PoP) where the VNF instance is placed. Intuitively, the amount of traffic flows that pass a VL instance should not exceed the VL capacity to avoid any overload at the VL instance. Likewise, the amount of incoming traffic flows to a VNF instance should not exceed the VNF processing capacity. (These constraints will be covered in the following paragraphs) Therefore, traffic flows for a network service (i.e., a VNF-FG) should be distributed to multiple NFPs depending on resource and capacity constraints for VNF and VL instances. Moreover, multiple network services can be supported by distributing traffic flows for each network service. Therefore, another decision variable, traffic flow ratio (TFR), is defined as: Traffic flow ratio (TFR): the ratio of the traffic flows distributed to each NFP. Therefore, the amount of traffic flows that passes on each NFP is the product of TFR and the amount of incoming traffic flows for a network service. Note that TFR and the amount of incoming traffic flows can be computed by measuring the amount of traffic flows that passes on each VL. The constraints regarding the amount of network traffic and capacity of VNF and VL instances can be specified as follows. Network traffic conservation constraints: In the VNF-FG system model, the amount of network traffic should be conserved within a VNF-FG. That is, 1) the amount of incoming network traffic to a VNF instance should be equal (more or less in case of packet manipulation) to the amount of outgoing network traffic from the VNF instance; and 2) the amount of incoming network traffic to a VNF instance should not exceed the flow rate of the corresponding NFP which can be determined by TFR. Network traffic processing order constraints: As defined in the VNF-FG, the network traffic can be processed by a VNF instance only after being processed by the preceding VNF instances along the NFP. Similarly, the incoming network traffic to a NFP should be firstly processed by the VNF instance which is located at the ingress point of the NFP; and the outgoing network traffic from a NFP should be the result of processing by every VNF instance in the order defined by the NFP. Link and processing capacity constraints: The amount of incoming network traffic to a VL instance should not exceed the given link capacity of the VL to avoid any congestion at the link. Likewise, the amount of incoming network traffic to a VNF instance should not exceed the processing capacity of the VNF. This system model can be exploited to obtain the optimal solutions of network resource (i.e., VNF and VL instances) placement for network resource usage, network service throughput, and so on. This optimization problem can be solved, for example with linear programming (LP), by defining different objective functions.

In the evaluation model, three objectives are considered including, but not limited to, 1) load balancing, 2) flow throughput maximization, and 3) energy efficiency.

For load balancing for a network service, the remaining capacity for VNF instances and VL instances should be balanced to avoid any bottlenecks. To this end, the minimum remaining processing capacity for VNF instances and the minimum remaining link capacity for VL instances should be maximized.

On the other hand, the flow throughput considers both throughputs for VNF processing and for VL instance. Then, the throughput of an NFP can be calculated as the product of TFR and the sum of capacities, and the total throughput is the sum of computed throughputs for all NFPs. By maximizing the total flow throughput, it is possible to reduce the network service time.

Since each VNF instance consumes an amount of energy for processing its function and transmitting/receiving traffic flows across VL instances, the energy consumption for each VNF instance should be minimized for energy efficiency of network services. Detailed model is under construction.

To support the aforementioned use cases, it is required to support resource management capability which provides service chain (or NFP) construction and adaptation by considering resource state or constraints of VNF instances and VL instances which connect them. The resource management operations for service chain construction and adaptation can be divided into several sub-actions: Locate VNF instances Evaluate the performance of VNF instances and VL instances Relocate (or scale) VNF instances to update a NFP instance Monitor state or resource constraints of a VNF instance, VL instances including NFVI-PoP interconnects As listed above, VNF instances are relocated according to monitoring or evaluation results of performance metrics of the VNF instances and VL instances. Studies about evaluation methodologies and performance metrics for VNF instances and NFVI resources can be found at . The performance metrics of VNF instances and VL instances specific to service chain construction and adaptation can be defined as follows: availability (or failure) of a VNF instance and a VL instance a topological location of a VNF instance CPU and memory utilization rate of a VNF instance a throughput of a VNF instance energy consumption of a VNF instance bandwidth of a VL instance packet loss of a VL instance latency of a VL instance delay variation of a VL instance The resource management functionality for dynamic service chain adaptation takes role of NFV orchestration with support of VNF manager (VNFM) and Virtualised Infrastructure Manager (VIM) in the NFV framework . Detailed functional building block and interfaces are still under study.

IETF SFC WG provides a new service deployment model that delivers the traffic along the predefined logical paths of service functions (SFs), called service function chains (SFCs) with no regard of network topologies or transport mechanisms. Basic concept of the service function chaining is similar to VNF-FG where a network service is composed of SFs and deployed by making traffic flows traversed instances of the SFs in a pre-defined order. There are several works in progress in IETF SFC WG for resource management of service chaining. defines SFC control plane that selects specific SFs for a requested SFC, either statically or dynamically but details are currently outside the scope of the document. There are other works which define the control plane functionality for service function chain construction and adaptation but details are still under study. While and provide detailed mechanisms of service chain adaptation, they focus only on resilience or fail-over of service function chains.

In SFC WG, describes requirements for conveying information between Service Function Chaining (SFC) control elements (including management components) and SFC functional elements. It also identifies a set of control interfaces to interact with SFC-aware elements to establish, maintain or recover service function chains.

|SFF |----->|SFF |------->|SFF | | | | Node |<----| |<-----| |<-------| | | | +-----------+ +----+ +----+ +----+ | | | | | | | |C2 ------- | | | | | | +-----------+ C4 | | V +----+ +----+ | SFC Proxy |--> | | | SF | |SF | +-----------+ | | +----+ +----+ | | |C3 |C3 | | SFC Data Plane Components V V | +-------------------------------------------------------------+ ]]>

The service chain adaptation addressed in this document may be integrated into the SFC Control & Management Planes and may use the C2 and C4 interfaces for monitoring or collecting the resource constraints of VNF instances, NFVI-PoP interconnects and VL instances. To prevent constant integration between the application and probing functions we would propose a 3-tier architecture per NFVI-PoP. Top level application control at the SFC Control & Management Planes An abstraction layer between the application layer and the probing layer. This would decouple NFVI and link monitoring methods from the application layer A probing layer that monitors VNF, physical and virtual link resources Note that SFC does not assume that Service Functions are virtualized. Thus, the parameters of resource constraints may differ, and it needs further study for integration.

TBD.

TBD.

In addition to the authors, the following individuals contributed to the content. Insun Jang Korea University zerantoss@korea.ac.kr

&rfc2119; ETSI ETSI ETSI ETSI ETSI

The authors would like to thank Sukjin Choo and Myeongsu Kim for the review and comments.