What is X-FIT?

X-FIT is a distributed time series analysis and forecasting system that uses an ensemble of models. The models compete in a distributed tournament, where the system performs cross-validation and selects the top models based on the mean absolute scaled error (MASE) on the out-of-sample forecasts during the cross validation. Then the final forecast is obtained by optimally combining the forecasts from the top performing models.

How X-FIT Forecasts

X-FIT is a highly scalable, application aware machine learning engine that powers Nutanix Prism capabilities such as Predictive Capacity Planning.

Unlike capacity planning tools that extrapolate historical data patterns, X-Fit is application-aware and employs multiple predictive algorithms that continually compete with one another to determine the most precise forecast.

How the Algorithm Works

The algorithm monitors predefined set of metrics daily and publishes new/adjusted baseline values for each metric. A number of factors go into determining the baseline values. These include seasonality and trending, as well as a historical analysis of 27 metrics covering VMs, hosts, and clusters for the previous 21 days. 

For each metric, the algorithm fetches past 21 days of data and publishes predictions for the next 7 days. It looks for:

• Seasonality (e.g. Weekend CPU utilization is very low)
• Trend (e.g. Average CPU utilization is increasing) 

Where is Machine Learning Used?

• Actionable Capacity Forecasting
  • Forecast of CPU, memory, and storage capacity runway (time remaining)
• Based on machine-learned consumption behavior from running workloads
• Detail trend of hosts (in CPU and memory) and containers (in storage) level
• Recommendation of appropriate scale out options (node types and numbers)