• United States
John Edwards
Contributing writer

How AI can improve network capacity planning

Feb 05, 20199 mins
Network Management SoftwareNetworking

The tricky art of provisioning network bandwidth is getting help from new tools. AI and machine learning provide the ability to precisely predict network needs.

abstract binary man ai matrix by ankarb getty
Credit: Ankarb / Getty

Network capacity planning aims to ensure that sufficient bandwidth is provisioned, allowing network SLA targets, such as delay, jitter, loss, and availability, to be reliably met. It’s a complex, error-prone task with serious financial implications. Until recently, the network data necessary for insightful capacity planning was generally only available via static, historical, after-the-fact reports. This situation is now rapidly changing.

“By pairing advanced data science and cognitive technology such as AI and machine learning, IT can drive new and smarter predictive insights to improve network capacity-planning accuracy,” says Ashish Verma, a Deloitte Consulting managing director specializing in cognitive analytics. “This helps organizations unleash data to make more agile decisions, improve operational wisdom, avoid downtime and create a better user experience.”

Although AI-supported network capacity planning is still in a very early stage, most capacity planning vendors, including large and small players like Cisco, NetBrain, Aria Networks, Flowmon and SolarWinds, are already beginning to incorporate some form of AI technology into their offerings, or plan to do so in the near future. Meanwhile, AI technology vendors, such as IBM Watson, are also looking to enter the game.

AI bolsters traditional network monitoring

Leveraging AI to analyze data from multiple sources offers much greater accuracy than the traditional network monitoring tools that look strictly at link utilization, notes Fredrik Lindstrom, a manager in the U.S. CIO Advisory practice of business consulting firm KPMG. “AI also enables modeling of different performance scenarios and ties network performance to application performance to determine how applications are impacted in different performance scenarios.”

AI-driven machine learning applied to network performance allows a network controller to learn from experience while it enhances the network.

“As it learns, the analytics model it uses for decisions is optimized and becomes a better representation of the true intent of the network and its business goals,” says Duval Yeager, analytics and machine-learning subject matter expert at Cisco. “This provides accurate capacity planning as the network grows, changes, and applications and users are added – both locally and in the cloud.”

AI and machine learning methods can be effectively applied to traffic prediction/forecasting, traffic pattern detection, online learning and automated decision making, observes Yan Huang, assistant professor of business technologies at Carnegie Mellon University’s Tepper School of Business.

“Advanced machine-learning algorithms can take large-scale and highly granular network data as inputs to generate precise demand forecasts for each node in the network and detect intertemporal patterns/trends in network traffic and utilization,” Huang explains. “The improved traffic and demand prediction will enable more accurate assessment of network capacity requirements and reduce the need for resource over-provision.”

Early detection and discovery of intertemporal patterns or changes in network traffic allow organizations to take proactive actions to ensure network performance. “Sophisticated predictive models can be combined with optimization and/or simulation techniques to automatically generate the optimal network structure or structures and the corresponding capacity and resource plans,” Huang says. Such plans can then be tailored to the specific performance metrics the organization cares most about.

AI technology can also process real-time traffic data and make routing and allocation decisions dynamically, based on real-time network conditions. “It also enables on-demand models for incremental capacity provisioning,” Huang explains. All of these factors can significantly reduce capital expenditures and operating expenses related to network development, maintenance and refinement, while lowering the effort required from IT professionals to manage such activities.

Once installed and properly configured, network AI technology can automate network capacity planning, accounting for both the financial and risk appetite of the organization. “AI can analyze many different data points in real-time or near-real-time, which is critical as organizations migrate to virtualized network overlays across their data centers, cloud environments and WAN,” Lindstrom says.

AI can also be used to analyze network traffic patterns in various ways, helping organizations gain insights into exactly what’s running across the network as well as the overall network load.

“This detail is useful for short- and long-term capacity planning,” explains Doug Tamasanis, chief architect and senior director of networks and security for Kronos, a workforce management software and service provider.

In the short term, AI can predict daily traffic bursts at granular levels, such as application, location, technology and protocol. These findings can then be used to protect against performance degradation during peak periods. “In the long term, an AI system can perform optimal capacity planning, anticipating when short-term bursts can’t be met and [when] whole scale upgrades are required,” Tamasanis notes.

AI-powered capacity planning: Getting started

The best way to get started with AI-powered capacity planning technology is to acquire a proven technology that has already achieved some level of success and enterprise acceptance, suggests Marcel Shaw, federal systems engineer for Ivanti, an IT asset and service management software provider.

“Meanwhile, administrators should approach recommendations produced by AI learning algorithms with caution,” he says. “AI learning algorithms will drastically improve over the next several years, so it will be important for customers to be patient and allow AI technology to mature before placing all their trust into the capacity requirements recommended by the AI solution.”

Start small, in terms of data sources and monitoring scope, Lindstrom advises. “It’s critical that data sources are reliable and consistent and that the [AI] system is able to generate the baseline over at least one full business cycle,” he explains.

Deploying network port replication to key network devices is the best way to provide a data stream capable of feeding an analytical platform, Tamasanis notes. Specific systems, such as wireless controllers, VPN concentrators and firewalls, can be configured to stream data directly. “Any AI system will require these types of feeds, and the more coverage across the network the better,” Tamasanis says. “The key is to get the maximum volume of data to the [AI] platform.”

It’s also important to provide the right data in the proper context. “Prepare your data for easy ingestion into the solution and make sure it’s delivering the view of your network capacity that’s relevant to your goals,” says Murthy Garikiparthi, director of engineering at cloud platform developer OpsRamp. Once the data pipeline has been established, and speeds and feeds are aligned, the [AI] solution can begin monitoring data for specific behaviors. “Finally, once AI begins making recommendations, the IT operations team can set up policies for automation that act on these insights,” Garikiparthi suggests.

Tamasanis stresses the importance of selecting the right AI platform. “Some platforms will fit some companies better than others,” he notes. “This natural variation is both an appealing and detrimental feature of AI analytics.” Tamasanis also suggests shying away from automated configurations. “While appealing from a reaction time, misinterpretation of data could introduce performance-lowering effects,” he cautions.

Misconceptions about AI resources, accuracy 

Perhaps the biggest misconception regarding the use of AI for network capacity planning is that the technology isn’t especially resource hungry, particularly in terms of human interaction. This illusion is “exacerbated by some vendors, who give the impression that you can just install the tool, and it will do everything without anyone managing,” Lindstrom says.

Another misconception is that AI-based network-capacity planning is an all-or-nothing game. Enterprises should work with their vendor on a phased approach where they deploy solutions in a modular fashion and focus on use cases where value is the greatest, Yeager recommends. Such an approach is particularly important in light of the fact that many future network elements will be cloud-based and subscription dependent. “Deploying in a phased, use-case approach will guarantee that IT managers are not paying for cloud subscriptions on services and solutions that aren’t deployed yet,” Yeager explains.

The biggest misconception about using AI for capacity planning is that the AI solution will always be accurate, Shaw says. “Until AI solutions mature, it will be important for administrators to verify and question recommendations that have been provided by an AI-powered capacity-planning solution.”

Pitfalls of AI capacity planning 

Like any emerging technology, AI capacity planning comes with its share of landmines, ready to devastate the innocent and unwary. Eager adopters tend to go too big, too soon and expect immediate results without any system fine tuning, Lindstrom says. “If the system does not have enough data to analyze, or the data is not reliable or consistent, the tools will not generate an accurate picture of the network or the performance issues of the network,” he cautions.

It’s also important to steer clear of vendors offering an incomplete selection of products or services. “Most solutions on the market today that offer great advantages for one small part of the network don’t offer a complete network solution for wired, wireless, devices, clients, applications, security, policies, cross domain, WAN, cloud and data center,’ Yeager observes.

Using AI products and services that aren’t based on an open platform will make it difficult for intelligence-oriented services to spread to other parts of the organization. “Future operations will no longer be siloed, and departments will share network, services and operational and usage data in order to enhance all departments enterprise-wide,” Yeager says.

Since AI is a relatively new technology, adopters often encounter resistance from old-guard managers and staff, intent on protecting institutional knowledge, legacy workflows and their jobs. “The old way of doing things can’t be so rigid as to reject the benefits of AI.” Garikiparthi warns.

AI offers an opinion, not a conclusive statement, Tamasanis says. “Integrating such a tool with existing network equipment and engineers will be an evolutionary process.”