CPU Usage Spikes

[Vollkornaffe]

When importing a large (geometry) node tree, we see very intense CPU usage.

Since we're not doing anything multi-threaded, it must be some geometry node evaluation.

Maybe we can turn that off while we are importing.
It could significantly improve the runtime.

[Vollkornaffe]

This is slowing development because tests run longer.

Hiding things in the viewport doesn't help.

[Vollkornaffe]

Built Blender with RelWithDebInfo and these flags:

CMAKE_CXX_FLAGS_RELWITHDEBINFO   -O2 -g -DNDEBUG -fno-omit-frame-pointer -mno-omit-leaf-frame-pointer

Then ran the import of the bejewelled geometry nodes and started this:

PID=$(pgrep -n blender)
sudo perf record -F 199 -g --pid $PID -- sleep 20
sudo perf script > out.perf
stackcollapse-perf.pl out.perf > out.folded
flamegraph.pl out.folded > perf.svg

Resulting in this flamegraph:

[Vollkornaffe]

It doesn't really look like there's "anything happening", but rather like there's a lot of pressure on the scheduler?

vmstat 1 shows that the number of context switches jumps from around 6k to 35k.

[Vollkornaffe]

sudo perf top shows

  16.26%  [kernel]                              [k] update_curr                                                                                                                                                                                               
   8.84%  [kernel]                              [k] pick_task_fair                                                                                                                                                                                            
   5.75%  [kernel]                              [k] __schedule        

[Vollkornaffe]

Following a discussion on Discord: https://discord.com/channels/1336035911766511687/1440413408993345656/1463914253164482620

Setting up many nodes through Python is currently an O(n^2) problem

We confirmed that it's slow regardless of the number of threads.
One can start Blender with fewer threads: --threads 1.

There's still "something" going on with the other threads, but it's not affecting the import's runtime much.

One thing that could be done here is to profile the import in this restricted setting, but in the end, it'll require some significant changes on Blender's side to be fast.