We have added AI metrics to complete the suite of workload metrics available in the Power BI Premium capacity metrics app (version 18.104.22.168), making it a one stop shop for admins. The app supports monitoring the health of your premium capacities, providing you valuable information on how best to fine tune your workloads, and when to scale so that your users get the best possible experience.
Today, we want to highlight the AI metrics, which provide an easy and informative way to monitor and identify potential issues.
AI summary metrics
The dashboard shows the 7-day summary of all the capacities that you are an admin of. In the following example, a total of 685 AI functions were executed when dataflow refreshes were scheduled with a 100% success rate at max CPU of 84.38% and max memory usage of 1.87 GB.
AI function duration in the past 7 days took 11 mins at the most and 10 mins on an average. Max and average wait time in the capacity is 0 milliseconds. Wait time is the time between the scheduled execution and actual start time.
AI detailed metrics
We’ve added an additional tab to the report.
The app now provides detailed metrics on the health of the AI workloads within your capacity. The top chart shows the memory consumption by AI workloads. You can set the memory limit for the AI workload per capacity. When memory usage reaches the memory limit, you can consider increasing the memory limit or moving some workspaces to a different capacity.
On the AI page, the “Overall Usage” table shows the details of AI functions executed in the past 7 days : the total count, system reliability, average/max wait times, average/max duration in milliseconds and the total size in bytes. The app groups these metrics by workspace, then dataflow, then AI function.
The AI workload can execute as many concurrent requests as there are number of cores in the capacity. In a P1 capacity, 6 of the AI function will wait to start execution if say 10 AI function execution requests were made at the same time. Long wait times is a sign that a capacity is becoming busy. If AI functions are waiting, you should be able to identify the active dataflow in a time period, based on total request and duration using the “Overall Usage” table. High memory utilization in the AI workload can be inferred from the memory spikes in the memory consumption chart on the top of the AI page. You can identify the dataflow consuming most of AI memory during the time period by sorting on the total size ( sum of the input and output size in bytes in memory) . As a mitigation, you can move the workspace containing the expensive dataflow to a different capacity.
Be sure to submit your ideas for new features. Learn more about Power BI Premium and monitoring premium capacities. To install the Power BI Premium Capacity Metrics app, click here.