A slow report is fixed with one repeatable loop: measure with Performance Analyzer, isolate the query in DAX query view, find the engine-level cause with DAX Studio, fix it, and measure again.
Why it matters
Each visual sends its own DAX query. If you do not know whether time goes to the query, the drawing of the visual, or something else, you will tune the wrong thing. Microsoft's own advice is symptom, diagnosis, fix, verify.
Step by step: the tuning loop
- Locate the bottleneck. Clear the cache, open Performance Analyzer, start recording, and refresh the visuals. Compare the timings per visual.
- Isolate the cause. Copy the slow visual's DAX query into DAX query view. If it is fast there but slow in the report, the problem is rendering. If it is slow everywhere, the problem is the measure or the model.
- Go deeper if needed. Paste the query into DAX Studio, turn on Server Timings and look at the Query Plan.
- Fix one thing. Change one measure, one relationship or one visual.
- Measure again. Run Performance Analyzer again to confirm the gain. Repeat.
Read the Performance Analyzer pane
Open it from the Optimize ribbon in Desktop. In the service, open the report, select Edit, then View > Performance Analyzer.
| Timing | What it measures |
|---|---|
| DAX query | From the visual sending its query until the model returns results |
| Direct query | For DirectQuery tables, the time the source takes to answer |
| Visual display | Time for the visual to draw, including fetching web images or geocoding |
| Other | Preparing queries, waiting on other visuals, background work |
| Evaluated parameters (preview) | Time spent evaluating field parameters |
Buttons you will use: Start recording, Stop, Clear, Refresh visuals, Analyze this visual (reloads one visual), Copy query, Run in DAX query view (Desktop only) and Export (saves the pane as JSON).
Compare an old and a new measure side by side
DAX query view lets you define a trial measure just for the query, so the model stays untouched. The names below are placeholders and the query was not run:
DEFINE
MEASURE 'Sales'[Total Sales v2] =
SUMX ( 'Sales', 'Sales'[Quantity] * 'Sales'[Net Price] )
EVALUATE
SUMMARIZECOLUMNS (
'Date'[Year],
"Current", [Total Sales],
"Trial", [Total Sales v2]
)Check that both columns match before you compare their speed.
Common causes and what Microsoft suggests
| Cause | What to do |
|---|---|
| Too much on one page | Keep to about eight visuals per page. Move detail to drillthrough pages and tooltips. |
| Heavy visuals | A table with 20 measures builds a large query. Scatter charts with 100,000 points or tables with 50,000 rows render slowly. Reduce measures or add Top N filters. |
| Ambiguous relationships | Keep single active relationships and use USERELATIONSHIP for alternates. Use bidirectional filtering only when needed. |
| High-cardinality keys | GUID keys are slower to traverse. Prefer integer surrogate keys. |
| Wide filter removal | REMOVEFILTERS or ALL on large tables forces full-table work. Scope them narrowly. |
| Large many-to-many bridges | Consider role-playing dimensions or consolidated tables. |
| DirectQuery | Check that Power Query steps fold, that the source has indexes, and consider Import for summary data with DirectQuery for detail. |
If DAX time is high, look in the query for FILTER over large tables, iterators and repeated subexpressions.
Let a tool review the model
The Best Practice Analyzer checks a model against more than 60 rules covering performance, DAX, error prevention and maintenance. In Fabric, open the BPA sample notebook from your semantic model in the service (Home ribbon or Model health menu). A companion Memory Analyzer notebook shows storage statistics. Without Fabric, Tabular Editor offers the same kind of check.
Watch-outs
- Performance Analyzer durations include time spent queued, because most canvas work runs one step at a time on a single UI thread.
- Clearing the pane deletes its data for good. Export first.
- Queries are cached as you use a report, so clear the cache first or later runs will look faster than they are.
- The query a visual writes is more verbose than one you would write by hand. Copilot in DAX query view can simplify it, but its output may contain errors and needs checking.
- Run in DAX query view is not available in the service. Copy the query, choose open data model, and run it in DAX query view there.
- Background, not from the Microsoft pages: Server Timings split work into the storage engine (reads and compresses data, can use many cores) and the formula engine (runs your DAX logic, one core). A slow formula engine usually means a measure needs rewriting; a slow storage engine usually means too much data is scanned.
Best practices
- Change one thing at a time and record the before and after numbers.
- Run the Best Practice Analyzer after you build or heavily change a model, not only when something is slow.
- Reduce what a page asks for before you rewrite DAX.
- Keep DAX Studio for engine-level questions. Performance Analyzer and DAX query view handle most cases.