Download presentation
Presentation is loading. Please wait.
1
An overview of roadmap plans for the APSD
Big Data-Fueled Feedback Loops in SDN/NFV Matt Olson CenturyLink An overview of roadmap plans for the APSD Big Data Analytics Platform (“BDAP”). B=Big Data R= Repository A = Artificial Intelligence I = Integrated N = Network DANN - Data & Analytics Neural Network Cognitive Computing March 31, 2016 Strata+Hadoop World
2
Who am I? Part of a small team deploying focused on transformation of networks with Software Defined Networking and Network Function Virtualization Background in Network Performance Engineering and Analytics Accidental foray into Data Science motivated by the desire to solve problems in the network space Relative newcomer to the field, but with years of experience data wrangling and building tools in Splunk, and a variety of other platforms
3
What’s my problem? CenturyLink is deploying a cloud based SDN/NFV platform capable of transforming network service delivery. The SDN/NFV platform is awesome for rapid and automated provisioning, but . . . To be really awesome, it needs a nervous system and brain Also, it’s crazy complicated - we need systems to help people manage the complexity until the robots are ready The challenge is to build a brain for the SDN/NFV fabric. At its core, this is a Big Data and an algorithmic challenge.
4
Why does it matter? SDN is poised to transform Telecom completely
“Within 5 years SDN/NFV will be the mainstream option for service providers deploying cloud and network architectures “ Tech Target, February 2015 CenturyLink is leading that charge with a huge commitment “(CenturyLink) plans to have full global virtualization of its IP core network and data centers by ” FierceTelecom, October 22, 2015 Telecom is changing, and to be successful, we need to jump ahead of the curve and drive that change. Under the hood, SDN/NFV is highly complex – Success depends inherently on automation, and intelligent use of the data generated from the platform.
5
A little background - SDN
Software Defined Networking (SDN) is a natural corollary to service virtualization The lines between Network and Services are blurring – Services are networked, and network provisioning should be as agile as the service provisioning. This need has prompted the development of SDN API driven routing and switching Automated configuration and management of network connectivity Separation of the control plane and forwarding plane, allowing for use of commodity hardware Provisioning in minutes, rather than days! Scaling and cost advantages (OPEX and CAPEX)
6
SDN Framework Software! Programmable via APIs
Opportunity to use commodity hardware Centralized interfaces
7
A little background - NFV
Network Function Virtualization allows us to leverage cloud services to provision network functions in a dynamic and flexible manner Network services run on VM’s. Control and data plane elements are fully abstracted, leveraging virtualized compute resources for scaling, resiliency, and agile provisioning across the VM platform. Introduces sophisticated Layer functionality in areas such as firewalls, load balancers, and video and IMS services. API driven provisioning of complex ‘service chains’ across the Layer 2 and 3 fabric. Single source of state and transactional data related to services end-to-end and state of the network and elements delivering those services.
8
NFV Framework Heterogeneity & Complexity Homogeneity & Simplicity
Non-NFV Fragmented non-commodity hardware Physical install per appliance per site Hardware development large barrier to entry for new vendors, constraining innovation & competition Heterogeneity & Complexity Virtualize the Network… NFV – Network Functions Virtualization. SDN – Software Defined Network. Open source API driven At Cloud Scale… Scale-out, not up Commodity hardware + Acceleration Consumable REST API’s. (Automation / Reduce or No-touch) Value the data… Collect network data into a Data Lake Easy tools to access & mine the data. Homogeneity & Simplicity
9
Programmable Services Backbone
7 Countries 44 Total Global Locations Data Center & Telecom Facilities 1 University & 4 DoD Labs SDN NFV TeraPoP/Core PSB SDN/NFV MIA ORL XXX Orchestration & Automation API’s CTL Cloud MPLS Central Offices & Data Centers Donation of 22 years of computing time in less than 2 years (Cancer, Ebola, AIDS, Genome Markers) Distributed-NFV CPE w/X86 Virtualization & IoT Premise Based
10
PSB Stack for SDN/NFV Network as a Service Platform
API Orchestration NFV) Programmable Services Backbone Networking (SDN) Virtualization (OpenStack / VMWare) Hardware
11
Orchestration & Controllers
BRAIN and Adaptive Service Chains SVC2 SVC1 SVC3 Services Layer SVC4 SVCn API Portal OSS/BSS SDN/NFV Orchestration & Controllers TOSCA Engine Network Forwarding Graph BRAIN Prescribe Predict Inform Simple on the outside (service chain, and API wrapper, but this hides a tremendous amount of complexity) BRAIN - Big data Repository and Artificial intelligence Integrated Networking SDN/NFV Platform Generates massive amounts of metadata related to the platform, and the services provisioned on the platform The SDN/NFV platform is also highly complex, with multiple layers of abstraction, complex end-to-end service chains, and dynamic topology This presents a huge operational challenge! But . . . The platform provides an integrated source of rich Real Time data The platform provides API interfaces and inherent agility => A huge opportunity for automation and Intelligent services! Data Warehouses & Data Lakes
12
Big Data? Diverse metadata sources:
Platform health and state data (bare metal, hypervisors, VMs, Applications) Network interface statistics and network traffic data (packet stats, NetFlow, SFlow etc.) Service health and usage data (transactional data, active test and probe data, QoS) Reference and topology data – Needed to enrich and make sense of platform and service data
13
Machine Learning and Analytics
The availability of rich platform data and service level data allows the application of Machine Learning techniques, and the training of models to drive optimized service chain architecture. Supervised learning focused on measured service quality and availability Tuning for specific services offered (e.g. Voice, Video, etc.) Incorporation of service and node interaction and usage patterns / node interactions Incorporation of platform element and link failure modes Train models on accumulated historical data, and apply model derived analytics to real time streaming data for pattern detection.
14
Feedback Loops With trained models feeding on streaming platform data . . . Automated . . . Trigger detection (e.g. a surge in traffic between nodes, service degradation observed on a network segment, etc.) Determination of remediation / optimization action Execution of orchestration action via API calls Evaluation of results with ‘commit’ or roll-back Through the use of Neural Network modeling and multi-dimensional feedback loops, network and service rebalancing can be supported in a broad manner enabling network and service resiliency, even under conditions of highly dynamic load shifts and/or multiple failures. Predictive, as well as responsive – Anticipate emerging needs Add contextual awareness and service intelligence
15
Use Cases - Adaptation
16
Target Architecture Streaming and Batch Ingest
Connectivity to various reference data stores (RDBMS and API interfaces) Persistence in HDFS with batch analytics using MapReduce Spark stack for analytics MLlib – Machine Learning TensorFlow - Deep Learning
17
Reference and Topology
BRAIN PSB Streaming Data Apex / Kafka Hadoop People Data Science Reference and Topology Spark R / SQL MLlib Robots
18
Practical Considerations - Scaling and Leverage
We are a very small team, ‘the tip of the spear’ So We need tools which are solid, resilient and scalable. Ideally, we need pre-existing solutions for scaling, resiliency, governance, etc. so we can focus our energies on new development further up the stack Ingest: DataTorrent (Apache Apex) Persistence: MapR (Hadoop and Mapr-FS) Presentation and analytics: Splunk / Hunk
19
More Practical Considerations – Legacy Integration
This is an interesting technical problem, but the real issue is that we’re brownfield . . . Interconnected legacy networks Legacy organizations Legacy systems Scarcity of Big Data skills Incremental incorporation of automation. Human interaction and exploration in a familiar environment is key Integration with existing legacy OSS platforms is a must => Splunk (with Hunk)
20
Splunk/Hunk Architecture
Analyze Explore Visualize Dashboards Share Hadoop Client Libraries Hunk offers Full-featured Analytics in an Integrated Platform Explore, analyze and visualize data, create dashboards and share reports from one integrated platform. Hunk enables everyone in your organization to unlock the business value of data locked in Hadoop Hunk integrates the processes of data exploration, analysis and visualization into a single, fluid user experience designed to drive rapid insights from your big data in Hadoop. Enable powerful analytics for everyone with Splunk’s Data Models and the Pivot interface, first released in Splunk Enterprise 6. And Hunk works with what you have today Hunk works on Apache Hadoop and most major distributions, including those from Cloudera, Hortonworks, IBM, MapR and Pivotal, with support for both first-generation MapReduce and YARN (Yet Another Resource Negotiator, the technical acronym for 2nd generation MapReduce) and Amazon EMR and S#. Preview results and interactively search across one or more Hadoop clusters, including from different distribution vendors. Customers can also use the ODBC driver to feed data from Hunk to third-party data visualization tools or business intelligence software. Because it leverages our experience across thousands of organizations, we’ve naturally made Hunk easy to deploy. Hadoop Clusters
21
Splunk / Hunk Unified Search
Intelligently Search Across Real-Time and Historical Data Using the Same Splunk Interface Intelligently search across real-time and historical data with unified search. With one simple search you can access data from Splunk Enterprise and Hadoop. Optimize Splunk Enterprise search head performance for real-time monitoring, alerting and dashboarding with short-term historical context Hunk search, analyze and visualize months or years of historical data in Hadoop Run unified queries and dashboards across Splunk Enterprise and Hunk Real-Time Data Historical Data in Hadoop
22
Splunk / Hunk – Legacy Integration
Immediate support for OSS / Operational support functions Existing REST API interface to Northbound systems Support for streaming ingest as well as Hadoop data store and batch processes Leveraging existing resources and expertise Ease of use – Coping with scarcity of Specialized Big Data related skills Splunkers and Splunk community Existing infrastructure and resources Spanning all relevant data sets Legacy Network data (already streaming into Splunk) – Across all relevant networks, bridging the legacy network and Virtualized RDBMS (via DBConnect) – reference, Topology, and customer data Native Splunk Indexed data Hadoop via HUNK
23
Splunk / Hunk – User Interface
Opportunity for exploration of the data For developers – Rapid development with low barrier to entry For Operations and Engineering Search language, Data Models, and Pivot Interactive / Real Time Data Drill Down Flexible Schema on Search (with Hunk support for HIVE Metastore) Visualization and analytics Operational support Data Governance RBAC Support for LDAP Single Sign on Resource bundling into Apps Environment for development and sharing of views and analytics
24
Hunk Deployment Overview
Hunk shares all architectural aspects in common with Splunk Enterprise – it runs splunkweb and splunkd. It only differs operationally by its license key. Hunk NameNode (CLDB) SplunkWeb ResourceManager Hadoop Cluster Splunkd ERP DataNode 1 NodeManager Linux Local HDFS The Hunk license will enable Hunk ERPs to interface with the Hadoop framework. Splunkd on Hunk determines if a search addressed is data external to Splunk index and dispatches a job accordingly. Your data DataNode n Linux Local NodeManager
25
Mixed-mode Search Previews data and allows users to search interactively by pausing and refining queries Pause or stop MapReduce jobs Preview results Switch over time preview Splunk Stream Hunk starts the streaming and reporting modes concurrently. Streaming results show until the reporting results come in. Allows users to search interactively by pausing and refining queries. This is a major, unique advantage of Hunk compared to alternative approaches such as Hive or SQL on Hadoop which require fixed schema in an effort to speed up searches, while Hunk retains the combination of schema on the fly with results preview. - Streaming mode (scanning) – this low-latency ERP process does the rudimentary search by retrieving a superset of the results from external stores. - Reporting mode – this ERP process runs the full Splunk remote search pipeline by pushing down the computation to DataNodes. This has the inherent advantage of using distributed computation such as Hadoop. A knowledge bundle is also pushed to HDFS and gets used for data enrichment. preview Hadoop MR / Splunk Index Time
26
BRAIN Integrated Analytics Platform
PSB Streaming Data Apex / Kafka / Splunk Fwd Splunk Index Splunk (SH) OSS Legacy Network Streaming Data People - Ops Dev BI Data Science Hadoop Hunk Reference and Topology Spark R / SQL MLlib Robots
27
Summary SDN/NFV offers tremendous transformative potential
The convergence of Big Data Technologies, ML, and SDN/NFV presents a huge opportunity to develop next generation service delivery But, introducing this transformation in an existing large network environment requires incremental change, and iterative improvement Need to Crawl -> Walk -> Run and work with tools that are resilient, scalable, open, and extensible Though the ‘Singularity’ is the ultimate goal, need to engage existing teams and systems with the data And, ultimately, the biggest challenges are organizational, and cultural, not technical . . .
Similar presentations
© 2025 SlidePlayer.com Inc.
All rights reserved.