gh-156860: Fix exponential asyncio.print_call_graph() output

[deadlovelll]

Fix exponential asyncio.print_call_graph() output

For more details see gh-156860

[brittanyrey]

Do you have a benchmark + perf outputs to demonstrate that the actual impact matches the expected impact laid out in the issue?

[deadlovelll]

Do you have a benchmark + perf outputs to demonstrate that the actual impact matches the expected impact laid out in the issue?

Hi! I'll send it within the day

[deadlovelll]

@brittanyrey hi again! Here it is

import asyncio
import pyperf

async def waits_for(*deps):
    await asyncio.gather(*deps)

async def build(shape, ntasks):
    fut = asyncio.Future()
    match shape:
        case "chain":
            cur = fut
            for _ in range(ntasks):
                cur = asyncio.ensure_future(waits_for(cur))
        case "dag":
            layer = [fut]
            for _ in range(ntasks // 2):
                layer = [asyncio.ensure_future(waits_for(*layer)) for _ in range(2)]
    await asyncio.sleep(0)
    return fut

def bench(loops, fut):
    t0 = pyperf.perf_counter()
    for _ in range(loops):
        asyncio.format_call_graph(fut)
    return pyperf.perf_counter() - t0

ssz = {
    "chain": (10, 20, 30, 40, 50),
    "dag": (4, 8, 12, 16, 20),
}

if __name__ == "__main__":
    runner = pyperf.Runner()
    for shape, sizes in ssz.items():
        for ntasks in sizes:
            loop = asyncio.new_event_loop()
            fut = loop.run_until_complete(build(shape, ntasks))
            runner.bench_time_func(f"{shape}-{ntasks}", bench, fut)

+----------------+---------+------------------------+
| Benchmark      | before  | after                  |
+================+=========+========================+
| chain-10       | 21.7 us | 22.5 us: 1.04x slower  |
+----------------+---------+------------------------+
| chain-20       | 42.3 us | 44.3 us: 1.05x slower  |
+----------------+---------+------------------------+
| chain-30       | 64.4 us | 67.6 us: 1.05x slower  |
+----------------+---------+------------------------+
| chain-40       | 85.6 us | 89.5 us: 1.05x slower  |
+----------------+---------+------------------------+
| chain-50       | 108 us  | 113 us: 1.04x slower   |
+----------------+---------+------------------------+
| dag-4          | 12.7 us | 13.0 us: 1.03x slower  |
+----------------+---------+------------------------+
| dag-8          | 58.7 us | 29.2 us: 2.01x faster  |
+----------------+---------+------------------------+
| dag-12         | 242 us  | 45.1 us: 5.37x faster  |
+----------------+---------+------------------------+
| dag-16         | 974 us  | 61.4 us: 15.87x faster |
+----------------+---------+------------------------+
| dag-20         | 4.21 ms | 78.1 us: 53.85x faster |
+----------------+---------+------------------------+
| Geometric mean | (ref)   | 2.43x faster           |
+----------------+---------+------------------------+

The regression 3-5% not in the noise range, I thought about not using id() for memo, and got the results:

+----------------+---------+------------------------+
| Benchmark      | before  | after                  |
+================+=========+========================+
| chain-10       | 21.6 us | 22.0 us: 1.02x slower  |
+----------------+---------+------------------------+
| chain-20       | 42.4 us | 43.3 us: 1.02x slower  |
+----------------+---------+------------------------+
| chain-30       | 64.3 us | 65.5 us: 1.02x slower  |
+----------------+---------+------------------------+
| chain-40       | 85.9 us | 87.1 us: 1.01x slower  |
+----------------+---------+------------------------+
| chain-50       | 108 us  | 110 us: 1.01x slower   |
+----------------+---------+------------------------+
| dag-4          | 12.7 us | 12.8 us: 1.01x slower  |
+----------------+---------+------------------------+
| dag-8          | 58.9 us | 28.7 us: 2.05x faster  |
+----------------+---------+------------------------+
| dag-12         | 242 us  | 44.2 us: 5.48x faster  |
+----------------+---------+------------------------+
| dag-16         | 972 us  | 60.0 us: 16.20x faster |
+----------------+---------+------------------------+
| dag-20         | 4.21 ms | 76.5 us: 55.10x faster |
+----------------+---------+------------------------+
| Geometric mean | (ref)   | 2.49x faster           |
+----------------+---------+------------------------+

Now it's 1-2%, and it doesn't grow with size

Thank you for the direction! I already pushed that