Follow Datanami:
January 20, 2020

Optimizing Utilization Forecasting with Artificial Intelligence and Machine Learning

Rebecca D'souza


Contextual blindness is one major risk that IT teams dealing with huge volumes of data are either unaware of or overlook, assuming it to be “the nature of their work.” Most IT teams use multiple tools for network analytics, which can cause the data obtained from these tools to have many missing links due to different formats or names for the same metrics. Without contextual data, it’s difficult to model a set of IT infrastructure resources, examine their dependencies, understand the overall business impact, analyze based on priority, troubleshoot, and resolve issues that crop up.

What IT team wouldn’t want a crystal ball that could predict the future of their organization’s IT infrastructure, letting them fix application and infrastructure performance problems before they arise?

Well, the current shortage of crystal balls makes the union of artificial intelligence (AI), machine learning (ML), and utilization forecasting the next best thing to anticipate and avoid issues that threaten the overall health and performance of your IT infrastructure components.

Utilization forecasting is a technique that applies ML algorithms to produce daily usage forecasts for all utilization across CPUs, physical and virtual servers, disks, storage, bandwidth, and other network elements, enabling networking teams to manage resources proactively. This technique helps IT engineers and network administrators prevent downtime caused by overutilization.

AI/ML-Driven Forecasting for IT infrastructure

An AI/ML-driven forecasting solution produces intelligent and reliable reports, taking advantage of the current availability of ample historic records and high-performance computing algorithms.


Without AI and ML, utilization forecasting relies on reactive monitoring. You set predefined thresholds for given metrics such as uptime, resource utilization, and network bandwidth, as well as hardware metrics like fan speed and device temperature. When a threshold is exceeded, an alert is issued.

However, this reactive approach will not detect the anomalies that happen below the set threshold, causing indirect issues that could be detrimental to your network’s health to go undetected. Moreover, traditional utilization forecasting can’t tell you when you will need to upgrade your infrastructure based on current trends.

How AI/ML-Driven Utilization Forecasts Are Better

A closer look at predictive technologies reveals the fundamental difference between proactive and reactive forecasting.

Without AI and ML, utilization forecasting uses linear regression models to extrapolate and make predictions based on existing data. This method involves no consideration of newly allocated memory or anomalies in utilization patterns. Also, pattern recognition is a foreign concept. Although useful, linear regression models do not give IT admins complete visibility.

Utilization forecasting driven by AI and ML, on the other hand, utilizes seasonal-trend decomposition using the Loess (STL) method. STL allows you to study the propagation and degradation of memory, as well as analyze pattern matching, whereby periodic configuration changes in a metric will automatically be adjusted. In simple terms, STL dramatically improves the accuracy of AI/ML-driven forecasting thanks to those dynamic, automated adjustments. Moreover, if any new memory is allocated, or if the memory size is increased or decreased for a device, the prediction will change accordingly; this is not possible with linear regression.

To forecast utilization proactively, you need accurate algorithms that can analyze usage patterns and detect anomalies in daily usage trends—without generating any false positives. Let’s examine a simple use case.

With AI/ML-driven utilization forecasting, you can find a minor increase in your office bandwidth usage during the World Series, the FIFA World Cup, or other sporting events. Even if you have a huge amount of unused internet bandwidth, this anomalous usage can be detected.

Likewise, proactive utilization forecasting lets you know when to upgrade your infrastructure by factoring your business’ new recruitment and attrition rates. With AI aiding intent-based networking, the IT admin is the gamemaster as well as the umpire, well-equipped with an extra pair of artificially intelligent eyes, ready to spot violations. Fueled by AI-based forecasting techniques, an organization’s mean time to repair (MTTR) is also bound to reduce.

Data-Driven Decision-Making with AI/ML-Driven Utilization Forecasts

Incorporating AI-based forecasting consequently means fewer bottlenecks and increased productivity for IT admins. While previously there was a need for constant, manual network monitoring, AI and ML can now be used to generate a forecast report so an IT admin has a clearer picture of the usage levels of various devices in the network. AI can provide predictions that play a pivotal role in an IT admin’s evaluation process, thus eliminating possible human error or biases. In other words, ML coupled with human intervention is possibly the best way to go about making all important, resource-dependent IT decisions.

IT admins can get a helping hand from an AIOps-enabled network performance monitoring (NPM) tool that offers data-driven recommendations in terms of consumable data formats of only useful information. This results in less experience-based problem-solving and more data-influenced problem avoidance. An AIOps-enabled NPM solution can alert the IT admin in advance about possible performance degradation. AI/ML-based forecasting enables organizations to raise automated device procurement requests well in advance, plan their internet billing and IT infrastructure budget, know if they need to reduce their internet usage, and modify the allocation and reallocation of virtual memory as required. This essentially means you can roll out the solution to a networking issue before it even knocks at your data center’s door.

Beyond forecasting, ML can also be used to improve anomaly detection. Here, adaptive thresholds for different metrics are established using ML and analysis of historic data, revealing anomalies and triggering appropriate alerts. Furthermore, application and infrastructure monitoring functions will also be improved when enhanced with AI and ML technologies.

Perhaps in the not-too-distant future, AI/ML-driven forecasting and monitoring will rival the predictive powers of the fabled crystal ball.

About the author: Rebecca D’souza is a product consultant at ManageEngine, a division of Zoho. Rebecca has a degree in electronics and communications engineering from the NIE Institute of Technology.

Related Items:

Can AIOps Save IT Management?

AIOps and the Kubernetes Revolution

Rethinking Data for AIOps-Driven IT Operations