Running Monte Carlo Pi with Flame App (Python)

By Klaus Ma
Flame Quant

Motivation

Monte Carlo estimation splits naturally into independent batches. Flame’s flamepy.app API packages the local Python project, registers a temporary application, runs decorated services on remote executors, and cleans up when the application is destroyed.

Pi example

The current examples/pi/python/main.py initializes an application before declaring the service:

import math
import numpy as np
import flamepy.app as app

app.init("pi-example")


@app.service()
def estimate_batch(num_samples: int) -> int:
    x = np.random.rand(num_samples)
    y = np.random.rand(num_samples)
    return int(np.sum(x * x + y * y <= 1.0))


num_batches = 10
samples_per_batch = 1_000_000
try:
    futures = [estimate_batch(samples_per_batch) for _ in range(num_batches)]
    insides = app.get(futures)
finally:
    app.destroy()

pi_estimate = 4.0 * sum(insides) / (num_batches * samples_per_batch)
print(f"Estimated PI: {pi_estimate:.10f}; actual PI: {math.pi:.10f}")

Calls to estimate_batch return futures. app.get waits for the batch and returns the counts. app.destroy releases the temporary application even if one of the calls fails. A future can also be passed into another service call without first retrieving its value locally.

Run the complete example from examples/pi/python in a Flame console with a running cluster:

uv run main.py

The estimate varies with random sampling. More batches increase total samples and provide more opportunities for parallel execution when executors are available.

References

Related Posts