Genesis 0.3.7 Headless Simulation: Periodic Time Spikes Every 10 Steps

[StoneWang-cn]

When running headless simulations with Genesis 0.3.7 (with the backend set to GPU), we observed significant time spikes occurring periodically at steps 10, 20, 30, ...—an issue that was not present in version 0.2.1.

Code as follows:

import genesis as gs
import time

gs.init(logging_level="warning",backend=gs.gpu)

scene = gs.Scene(show_viewer=False,)

plane = scene.add_entity(
    gs.morphs.Plane(),
)
franka = scene.add_entity(
    gs.morphs.MJCF(file="xml/franka_emika_panda/panda.xml"),
)

scene.build()
# Key: Add scene.step() time cost statistics
print(f"Starting scene.step() time cost test (total 100 steps)...")
total_step_time = 0.0

# Test 100 steps and calculate the time cost per step and average time cost
for i in range(100):
    start_time = time.time()
    scene.step()  # Timing core: Only measure the step itself
    step_time = (time.time() - start_time) * 1000  # Convert to milliseconds
    total_step_time += step_time
    print(f"Step {i+1} time cost: {step_time:.4f} ms")

Log as follows:

Step 1 time cost: 0.6731 ms
Step 2 time cost: 0.4427 ms
Step 3 time cost: 0.3746 ms
Step 4 time cost: 0.3872 ms
Step 5 time cost: 0.3574 ms
Step 6 time cost: 0.3569 ms
Step 7 time cost: 0.3574 ms
Step 8 time cost: 0.3591 ms
Step 9 time cost: 0.3591 ms
Step 10 time cost: 8.3952 ms
Step 11 time cost: 0.5143 ms
Step 12 time cost: 0.4044 ms
Step 13 time cost: 0.3672 ms
Step 14 time cost: 0.3803 ms
Step 15 time cost: 0.3626 ms
Step 16 time cost: 0.3536 ms
Step 17 time cost: 0.3591 ms
Step 18 time cost: 0.3548 ms
Step 19 time cost: 0.3479 ms
Step 20 time cost: 3.3112 ms
Step 21 time cost: 0.3839 ms
Step 22 time cost: 0.3545 ms
Step 23 time cost: 0.3519 ms
Step 24 time cost: 0.3462 ms
Step 25 time cost: 0.3483 ms
Step 26 time cost: 0.3586 ms
Step 27 time cost: 0.3476 ms
Step 28 time cost: 0.3493 ms
Step 29 time cost: 0.3479 ms
Step 30 time cost: 3.5298 ms
Step 31 time cost: 0.3800 ms
Step 32 time cost: 0.3536 ms
Step 33 time cost: 0.3531 ms
Step 34 time cost: 0.3514 ms
Step 35 time cost: 0.3428 ms
Step 36 time cost: 0.3488 ms
Step 37 time cost: 0.3479 ms
Step 38 time cost: 0.3428 ms
Step 39 time cost: 0.3426 ms
Step 40 time cost: 3.4425 ms
Step 41 time cost: 0.3858 ms
Step 42 time cost: 0.3502 ms
Step 43 time cost: 0.3483 ms
Step 44 time cost: 0.3495 ms
Step 45 time cost: 0.3483 ms
Step 46 time cost: 0.3407 ms
Step 47 time cost: 0.3469 ms
Step 48 time cost: 0.3362 ms
Step 49 time cost: 0.3364 ms
Step 50 time cost: 2.9070 ms
Step 51 time cost: 0.4036 ms
Step 52 time cost: 0.3495 ms
Step 53 time cost: 0.3426 ms
Step 54 time cost: 0.3703 ms
Step 55 time cost: 0.3362 ms
Step 56 time cost: 0.3486 ms
Step 57 time cost: 0.3531 ms
Step 58 time cost: 0.3507 ms
Step 59 time cost: 0.3543 ms
Step 60 time cost: 4.9796 ms
Step 61 time cost: 0.3481 ms
Step 62 time cost: 0.3467 ms
Step 63 time cost: 0.3500 ms
Step 64 time cost: 0.3433 ms
Step 65 time cost: 0.3431 ms
Step 66 time cost: 0.3464 ms
Step 67 time cost: 0.3386 ms
Step 68 time cost: 0.3417 ms
Step 69 time cost: 0.3536 ms
Step 70 time cost: 4.6794 ms
Step 71 time cost: 0.4110 ms
Step 72 time cost: 0.3462 ms
Step 73 time cost: 0.3476 ms
Step 74 time cost: 0.3448 ms
Step 75 time cost: 0.3393 ms
Step 76 time cost: 0.3538 ms
Step 77 time cost: 0.3457 ms
Step 78 time cost: 0.3471 ms
Step 79 time cost: 0.3471 ms
Step 80 time cost: 5.2340 ms
Step 81 time cost: 0.3831 ms
Step 82 time cost: 0.3586 ms
Step 83 time cost: 0.3524 ms
Step 84 time cost: 0.3459 ms
Step 85 time cost: 0.3493 ms
Step 86 time cost: 0.3467 ms
Step 87 time cost: 0.3424 ms
Step 88 time cost: 0.3488 ms
Step 89 time cost: 0.3371 ms
Step 90 time cost: 3.3836 ms
Step 91 time cost: 0.3951 ms
Step 92 time cost: 0.3486 ms
Step 93 time cost: 0.3488 ms
Step 94 time cost: 0.3517 ms
Step 95 time cost: 0.3471 ms
Step 96 time cost: 0.3495 ms
Step 97 time cost: 0.3414 ms
Step 98 time cost: 0.3471 ms
Step 99 time cost: 0.3483 ms
Step 100 time cost: 3.8764 ms

[duburcqa]

This is a know phenomenon. it should not affect the overall speed. It is due to periodic GPU synchronisation for return code error checking. Hopefully it will be solved by #2011