NFV: A Dynamic, Multi-Layer Resource Optimization...
Transcript of NFV: A Dynamic, Multi-Layer Resource Optimization...
NFV: A Dynamic, Multi-Layer Resource Optimization
Challenge Manjari Asawa
Product Management, HPE
Agenda
• Why NFV
• Dynamic, Multi-layer Resource Optimization
• Implementation Approach
Early 2010s:
• Virtualization became common in IT with VmWare ESXi, kvm
• X86 advances allowed faster I/O and faster packet processing
• Rackspace & NASA launched open-source OpenStack cloud orchestrator
• Intel released software libraries and driver for fast packet processing (DPDK)
• Advances in Programmable networks (SDN)
NFV White Paper: Oct, 2012
• Operators’ dilemma– New service required procuring and installing new hardware
– Large variety of proprietary hardware
– Competition from OTT Players required fast new service rollouts
• New ETSI initiative launched– at&t, BT, CenturyLink, China Mobile, Colt, DT, KDDI, NTT, Orange, TI, Telefonica, Telstra & VZ
• Aim to leverage virtualization for– Use of standard servers and storage
– Efficient resource sharing and utilization
– Faster time to market
– Flexible and programmable operations
Vision for NFV*
*Source: ETSI White Paper https://portal.etsi.org/NFV/NFV_White_Paper.pdf
Evolution to Telco Cloud*
*Source: ETSI White Paper https://portal.etsi.org/NFV/NFV_White_Paper.pdf
Enterprise vs Telco Cloud
Limited by both CPU and I/O Performance
Enterprise Cloud Telco cloud
Limited by CPU Performance
Standard, out-of-box platform software
10G, Software only switching
Few Large DCs (Primary and DR IT DCs)
Many and Small VMs: referred as Cattles
40G, low latency and jitter requires hardware techniques such as SR-IOV, PCI-PT.
Software is often augmented with hardware
Many smaller and distributed service center: POPs
Few and Large VMs: Referred as Pets
Larger packet size for user applications Small packet size for network switching and voice app
ETSI NFV Reference Architecture*
*Source: ETSI White Paper
S-GW
LTE/Wi-Fi
LTE EPC (P0)
P-CSCF/SBC
PSTN/PLMN(TDM/SS7)
2G/3G Ckt Switched Core + SMSC
eNodeB
Other Services
e.g. RCS, Voice Mail
BGCF
Serving-GW
Untrusted Wi-Fi (P0)
Ml
Cx
Mj
Mr
Mw
Trusted Wi-Fi (P1)
ePDGTWAG
S/I-CSCF
P-GW
Regulatory
(PSAP, CALEA, location)
Mw
X1/X2/X3
HSS TAS
SCC-AS IM-SSF IP-SM-GW
Mr
E-CSCF
MRF
IPX/NNI
MGCF/SigGW/
MGW
TDM/SS7
CAMEL/MAP/Gd
ISC
IBCF/TrGW
PCRF
DRA
Charging
Ml/Mm/Mi/Mg
Mg
Mb
Rf
4G IMS
SIP VoIP
networkSh
RxISC
Mobile IMS (VoLTE) Architecture
HPE IMS VNFsHSS, MRF, OCCP (TAS,SCC-AP, IM-SSF, IP-
SM GW)
NFV Orchestrator
Field Integrated Apps
Compute Virtualization(KVM, vmWare)
NetworkVirtualization
Specialized HW Servers Storage NetworkingWAN
Network
NFVILayer
SDNController
VIM (Openstack)
InfrastructureManagement
Orchestration, OSS, BSS
Deploying IMS in Virtualized Architecture
Multi-layer, Dynamic, Resource Optimization
vIMS Service Components
Physical Resources
Physical Topology
Virtual Topology
Services (composite VMs)
Applications E.g. Mobile Registration, Call Origination
E.g. SBC, S-CSCF, TAS
E.g.: SBC VM1, SBC LB, SBC Route Selector, Overlay n/w
E.g.: Compute 1 CPU, compute 2 memory, NIC 3
E.g.: Physical Connectivity among resources
Orchestration
Policies
Performance Management
Life Cycle Management
Constraints
Service Chaining
Physical Data Center Architecture
Source: NFV: Report on SDN Usage in NFV Architectural Framework ( http://www.etsi.org/deliver/etsi_gs/NFV-EVE/001_099/005/01.01.01_60/gs_nfv-eve005v010101p.pdf)
NFV Platform Requirements• Orchestration:
– On-boarding, global resource management, authorization
• VNF Management:– Lifecycle management, configuration, event reporting
• Virtualized Infrastructure Management (VIM):– Control and management of the NFVI compute, storage, and network
• Service Performance:– E.g. mouth-to-ear delay should be less than 200 msec.
• Operational:– Minimize power consumption– Adoption to traffic conditions– High availability– Security, co-existence with existing platforms, software upgrades etc.
Hierarchical Controller Options*
Single SDN Controller for WAN Unique abstraction per client for WAN resources
Each client has direct access to WAN resource.
*Source: NFV: Report on SDN Usage in NFV Architectural Framework ( http://www.etsi.org/deliver/etsi_gs/NFV-EVE/001_099/005/01.01.01_60/gs_nfv-eve005v010101p.pdf)
Hierarchical, Distributed, Resource Allocation• Objective:
– Optimal resource allocation and scheduling at each layer.
• Environment:– Fast, Real-time action and response– Distributed, stochastic, dynamic information
• Approach:– Understand application requirements– Understand VNFs– Know infrastructure– Formalize scheduling problem accounting for
– Translation of application requirement to VNF and service function chains requirements– Efficient placement of VNFs onto physical infrastructure
Understand Application in terms of VNFs
• Understand Application Requirements
– User experience should be comparable to traditional networks.
– Decomposition of monolithic applications to smaller functions.
Source: ETSI GS NFV 002: NFV Architectural Framework. http://www.etsi.org/deliver/etsi_gs/nfv/001_099/002/01.01.01_60/gs_nfv002v010101p.pdf
• Critical Open Problems:
– Efficient placement and chaining of virtual network functions.
– Optimal decomposition point of services (tradeoffs between elasticity and delay)
– Description of NS, VNFs, their relationships, performance requirements (descriptors)
– More
ETSI E2E Network Service Graph Representation*
*Source: ETSI GS NFV 002: NFV Architectural Framework. http://www.etsi.org/deliver/etsi_gs/nfv/001_099/002/01.01.01_60/gs_nfv002v010101p.pdf
Understand VNFs
• VNFs are getting complex – Need to work with database, processes– Often requires high availability – Self-healing and auto-scaling properties
• Interdependencies– Dependencies between VNFs and to external network.
• Framework is needed to understand VNFs and interdepdencies– Dependency on operating conditions
– Need of dynamic characterization
– Definition of appropriate measurement metrics
– Use of Machine Learning to predict changes to traffic, SLAs etc.
VNF Performance with Different Configuration*
0
100
200
300
400
500
600
700
800
900
1000
Regi
stra
tion
Requ
ests
/sec
Clearwater Performance Characterization
bono1small-sprout1small-homestead1small
bono1medium-sprout1medium-homestead1medium
bono1large-sprout1large-homestead1large
bono1medium-sprout2small-homestead2small
bono1large-sprout2medium-homestead2medium
Clearwater
Bono
Ralf
Sprout
Homer Ellis
Homestead
* Source: NFV-VITAL: A framework for characterizing the performance of virtual network functions,Cao et.al., 1st IEEE NFVSDN Conference, San Francisco, November 2015
Know Your Infrastructure in Detail
• Compute resource parameters– vCPU, CPU partitioning, CPU model– Huge pages, NUMA support, – vSwitch capabilities and requirements – NICs - speeds and feeds of NICs capabilities such as SR-IOV
• Network Layout– Connectivity– Link bandwidth
• Storage– How much, where
• Location and connectivity among and within data centers
Performance Variation on Heterogeneous Servers*
4250
75 70
195
100
0
50
100
150
200
250
Snort Suricata
VNF
Capa
city
(K
pack
ets
per
sec.
)
VNF
VNF Performance for different Server Configurations
ServerConfig1(1.2Ghz) ServerConfig2(2.0Ghz) ServerConfig3(2,8Ghz)
“Not all servers are created equal”
Different servers exhibit different performance for different VNFs
– Server support for different virtualization and optimization knobs e.g. SR-IOV, CPU speed, NUMA etc.
– Varying VNF requirements e.g. data plane throughput, storage access rates etc.
* Source: NFV-VITAL: A framework for characterizing the performance of virtual network functions,Cao et.al., 1st IEEE NFVSDN Conference, San Francisco, November 2015
Once we understand applications,VNFs and infrastructure, how tomanage VNFs and Infrastructuremapping to meet applicationrequirements efficiently?
Formalize Options at Each Layer• Applications:
– How much to decompose and how
• NS: – How many VNFs, and forwarding graph among them.
• VNFs: – Scale –up or Scale-out– Determine the way it scales – Watch for noisy neighbors– Auto-scale, auto-migrate
• Compute– Processor types– NICs/Ports types and numbers
• Decision at one layer affects performance of the others
Optimally Allocate Resources
• Should understand application and constraints.
• Should be hierarchical with abstraction at every layer
• Should keep detailed inventory
• Should be contextual: – The right thing to do may not be the same
– Contexual information should be represented and available
• Should be dynamic with fast response time
Intent Driven Framework
• Translate application requirement to action– Method to describe intent
– Map intent to appropriate technology
• Things to consider during mapping– State of VNFs (application analytics: include prediction)
– Current state of infrastructure (performance analytics)
– Environment constraints
– Management constraints
– Mechanism to resolve conflicts
Implementation Approach
Open Source Approaches*
OSS / BSS
Open Source VNFs
NFV ISG
Open SourceStandards
*Source: OPNFV
OpenStack Architecture*
* Source: http://docs.openstack.org/mitaka/install-guide-rdo/common/get_started_conceptual_architecture.html
Nova Filter Scheduler*
* Source: Openstack configuration guide http://docs.openstack.org/mitaka/config-reference/compute/scheduler.html
Uses knowledge of VNF and infrastructure to filter out the computes.
Scheduler projects
• Nova scheduler – Simple and fast.
– Projects such as watcher, blazar will bring more intelligence and control.
– Other open-source projects : Mesos, Kubertenes
• IBM Platform Resource Scheduler: – Dynamic, intelligent, policy-driven resource scheduler
• HPE labs Stringer : – Accounts for heterogenous nodes, service chaining, integer programming.
• Research: – Network-Aware Round Robin (NARR)
– Elastic Edge (E2) framework
– More
Watcher Architecture*
*Source: Openstack wiki https://wiki.openstack.org/wiki/WatcherArchitecture
Some NFV Related Open-source ProjectsName Description
Tacker Generic VNF Manager (VNFM) and a NFV Orchestrator (NFVO) to deploy and operate Network Services and Virtual Network Functions (VNFs) on an NFV infrastructure
Openstack Blazar, OPNFV Promise
Resource reservation and management : user can request the resources of cloud environment to be provided (“leased”) to his/her project for a specific amount of time, immediately or in future.
Kingbird Centralized service for resource management across multiple OpenStack instances in a multi-regiondeployment, such as centralized quota management, centralized view for distributed virtual resources, global view of tenant address space, synchronization of ssh keys, images, flavors, security groups, etc.
Watcher Watcher provides a flexible and scalable complete resource optimization service. Include metric receiver, optimization processor and action plan applier using machine learning algorithms. Can run in advise only mode or active mode.
Congress Framework for governance and regulatory compliance across services (e.g. application, network, compute and storage) within a dynamic infrastructure. Use a high-level, general purpose, declarative language to describe which states of the cloud are in compliance and which are not, and allows enforcement.
Network Intent Composition (NIC)
Enables the controller to manage and direct network services and network resources based on describing the “Intent” for network behaviors and policies instead of describing how to provide different services.
Doctor (OPNFV) NFVI fault management and maintenance framework supporting high availability of the Network Services on top of the virtualized infrastructure.
Relationship Among Projects
Nova Neutron
Ceilometer Heat
Infrastructure
compute storagenetwork
KVM Ceph
Virtual compute
Virtual storage
Virtual network
VNF1 VNF2 VNF3
OVS
Tacker
NFVO
TOSCA NFV profile
Monasca
WatcherPromise
(Resource Reservation)Blazar
Prediction
Doctor
Copper
Congress NIC
Vitrage(Root cause
analysis)
OVSNFV
KVMNFV
VIM
Service Function Chaining
Policy
Monitoring
NetVirt OVSDB
NFVI
MANO
ODL controller
VNFs
OPNFV OPENSTACK ODL
Verizon SDN-NFV Architecture*
*Source: Verizon Network Infrastructure Planning: SDN-NFV Reference Architecturehttp://innovation.verizon.com/content/dam/vic/PDF/Verizon_SDN-NFV_Reference_Architecture.pdf
at&t ECOMP architecture*
*Source: ECOMP (Enhanced Control, Orchestration, Management & Policy) Architecture White Paper, http://about.att.com/content/dam/snrdocs/ecomp.pdf
Summary
• Dynamic and stochastic environment
• Applications are demanding
• Hierarchical control is a necessity
• End-to-end optimization framework is needed.
Happy Birthday, Demos