[mini] Improve documentation for nsys usage (#684)

When running with nsys, stalling was observed in running more
sophisticated runs.
They come from the cuda events. Disabling the tracing of the cuda events
fixes the problem, hence the documentation should reflect that.

Author: SeverinDiederichs

docs/user/profiling.md

@@ -47,16 +47,19 @@ The iteration range is half-open: `[start, stop)`. For example,
 iterations 50 through 149. Invalid or negative values are ignored and treated as
 the default `0`.
 
-Run `nsys` with the CUDA profiler API capture range:
+For profiling longer, multithreaded runs, such as hundreds of ttbar events, the CUDA-event tracing needs to be disabled, as it was observed create stalls in `cudaEventRecord`. Furthermore, writing periodically the buffered CUDA trace data can help with memory issues. This can be done with the following options:
 
 ```console
 nsys profile --capture-range=cudaProfilerApi --capture-range-end=stop \
-  --trace=cuda,nvtx --sample=none --cpuctxsw=none \
+  --trace=cuda,nvtx --cuda-event-trace=false --cuda-flush-interval=10000 \
+  --sample=none --cpuctxsw=none \
   --stats=true --export=sqlite --force-overwrite=true \
   --output adept_transport_profile \
   <application command>
 ```
 
+Note that older Nsight Systems versions do not support the options `cuda-event-trace` and `cuda-flush-interval`. In that case, it is advised to use a more recent Nsight Systems version (tested with 2025.5.2).
+
 Open the report with:
 
 ```console