#FAQ

The checks below apply to GPU, NPU, and HAMi configurations. Where an example uses a GPU metric or item, substitute the normalized metric and item from the corresponding resource page.

#Why is no accelerator usage shown?

For the resource page you configured, check the following in order:

1. The source exporter or HAMi scheduler metric returns the labels required by the selected resource page.
2. The normalized allocation metric returns data in Prometheus (alauda_gpu_allocated_count, alauda_npu_allocated_count, or alauda_hami_virtual_memory_allocated_bytes).
3. The collection ConfigMap is in cpaas-system on the accelerator cluster and has the cpaas.io/slark.collection.config: "true" label.
4. The collection item and query match the recording rule for the selected resource page.
5. The agent has reloaded the configuration and enough time has passed for a collection cycle.

#Why is there usage but no bill?

Check that the Cost Model is linked to the GPU cluster and contains a price for the exact billing item and method. Confirm that the display ConfigMap item names match the collection items exactly.

#Why are all records using the default price?

The model-specific override matches the billing label modelName. Confirm that:

• the recording rule creates label_modelName;
• the value is copied into the billing record as modelName;
• the Cost Model key is exactly modelName;
• the value matches case-sensitively.

Project-quota records use the default project price because the quota metric has no GPU model dimension.

#Why is the displayed data delayed?

The sample configuration evaluates usage every five minutes and aggregates it hourly. Allow at least one collection interval and the billing processing time before checking the console.

#Why did changing the price not change old bills?

Cost Model changes do not automatically recalculate historical bills. Treat the new price as applying to subsequently processed data unless a separate, bounded recalculation procedure is performed.

#Can DCGM_FI_DEV_GPU_UTIL be billed directly?

No. Its numeric value is a utilization percentage, not a GPU count. The sample uses the existence of labelled GPU series to derive one allocated GPU per UUID.