Client Sampling

Intuition

In cross-device FL (millions of phones) only a small fraction of clients participate in each round — those that are charging, idle and on Wi-Fi. Simulations model this by sampling a fraction C of clients per round.

Precise explanation

• Number per round: $m = \max(1, \operatorname{round}(C\cdot K))$ (McMahan et al. use $\max(\lceil CK\rceil, 1)$; we use round — a design choice — and test it).
• Sample without replacement and sort for readability.
• Reproducibility per round: seed the generator with (seed, round_idx) — np.random.default_rng([seed, round_idx]) — so round 7 always selects the same clients regardless of how many earlier rounds were simulated.
• Over many rounds each client participates with frequency ≈ m/K. Cross-silo FL (a few hospitals) often uses all clients every round (C = 1).

Worked example

K = 10, C = 0.3 → 3 clients per round, e.g. [1, 4, 8].

Common mistakes

• Sampling with replacement (the same client twice).
• m = 0 for tiny fractions.
• One global RNG whose state depends on the simulation history.

Why this matters

Partial participation is a key source of noise and bias in FL; FL-10 uses this sampler.

References

• Communication-Efficient Learning of Deep Networks from Decentralized Data — H. Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, Blaise Agüera y Arcas (S028)
• Advances and Open Problems in Federated Learning — Peter Kairouz, H. Brendan McMahan et al. (S032)