When running the quadruped from the examples, I noticed a CPU bottleneck. This was confirmed using the profiler, and also by setting the lowest rendering resolution and still getting the same fps.
For reference, I used the same container on my local machine with an RTX 3060 Ti and an AMD Ryzen 5800x CPU, and the performance was significantly better (~15 more fps) on the local machine for the same resolution.
I also had the issue on this local machine, but was able to mitigate it by setting the cpu governor to performance (was on powersave by default). But this cannot be done on the VM since it has to be changed on the host.
Is there a workaround to avoid this CPU bottleneck ?
Hi, although I am not the original poster, I can confirm that with Isaac Sim version 4.5.0 and a similar server platform (AMD 7763 CPU, docker, etc), this issue is still present.
Once the GUI of Isaac Sim is started, it creates way too many threads (256 in my case) and consumes 20% of the total 256 cores, which is a bit confused for me.
If I limit the FPS in the developer settings to 15 fps, the CPU usage goes down a bit. However, I think it’s just a walkaround than a perfect solution.
I would be glad if you developers are interested in this issue and I am willing to provide more info if it helps. Thanks!
This issue has been partially addressed. Starting with Isaac Sim 5.0, SimulationApp automatically caps thread counts to min(cpu_cores, 32) via the limit_cpu_threads config parameter (defaults to 32). This prevents the runaway thread spawning
on high-core-count server CPUs.
If you’re using SimulationApp from Python, this is already handled:
# Default: capped at 32 threads
app = SimulationApp({"headless": True})
# Or set a custom limit
app = SimulationApp({"headless": True, "limit_cpu_threads": 16})
However, if you’re launching via isaac-sim.sh or isaac-sim.headless.native.sh directly (which was your workflow), the .kit app files don’t include this cap. You’ll need to pass the flags manually: