Fault-Tolerant Aggregation: Flow-Updating Meets Mass-Distribution

Flow-Updating (FU) is a fault-tolerant technique that has proved to be efficient in practice for the distributed computation of aggregate functions in communication networks where individual processors do not have access to global information. Previous distributed aggregation protocols, based on repeated sharing of input values (or mass) among processors, sometimes called Mass-Distribution (MD) protocols, are not resilient to communication failures (or message loss) because such failures yield a loss of mass. In this paper, we present a protocol which we call Mass-Distribution with Flow-Updating (MDFU). We obtain MDFU by applying FU techniques to classic MD. We analyze the convergence time of MDFU showing that stochastic message loss produces low overhead. This is the first convergence proof of an FU-based algorithm. We evaluate MDFU experimentally, comparing it with previous MD and FU protocols, and verifying the behavior predicted by the analysis. Finally, given that MDFU incurs a fixed deviation proportional to the message-loss rate, we adjust the accuracy of MDFU heuristically in a new protocol called MDFU with Linear Prediction (MDFU-LP). The evaluation shows that both MDFU and MDFU-LP behave very well in practice, even under high rates of message loss and even changing the input values dynamically.

💡 Research Summary

The paper addresses the problem of distributed average computation in networks where nodes have no global knowledge and communication is unreliable. Classical mass‑distribution (MD) protocols, which repeatedly share a fraction of each node’s current estimate with its neighbors, lose “mass” when messages are dropped, causing incorrect convergence. Flow‑Updating (FU) techniques, on the other hand, keep the original input value at each node and only exchange cumulative “flows”, making them naturally resilient to loss, but they have lacked rigorous analysis.

The authors combine these two ideas into a new protocol called Mass‑Distribution with Flow‑Updating (MDFU). In MDFU each node stores: (i) its current estimate e_i, (ii) for every neighbor j the cumulative inbound flow F_in(j) and outbound flow F_out(j). At the beginning e_i is set to the node’s input v_i and each outbound flow is initialized to e_i/(2·D_ij), where D_ij is the maximum degree of the two endpoints. In every round a node sends its current outbound flow to each neighbor; if a message arrives the receiver updates the corresponding inbound flow, otherwise it keeps the last successfully received value. After the communication phase the node recomputes its estimate from scratch as

 e_i ← v_i + Σ_{j∈N_i} (F_in(j) – F_out(j))

and then updates each outbound flow by adding e_i/(2·D_ij). Thus, when no messages are lost the algorithm behaves exactly like a standard MD protocol, while under loss it still retains the original inputs and only the flow records become stale.

The paper provides the first convergence proof for an FU‑based algorithm. For the loss‑free case (loss probability f = 0) the transition matrix P of the protocol is doubly stochastic and can be interpreted as the transition matrix of a reversible Markov chain on the network graph. Using conductance Φ(G) of the underlying weighted graph, the authors show that after

 r_c = 2·ln(n/ξ) / Φ(G)^2

rounds the maximum relative error ε(r) ≤ ξ holds for any 0 < ξ < 1. This bound mirrors classic MD analyses but is derived without assuming mass conservation because MDFU recomputes estimates from the original inputs each round.

When messages are lost independently with probability f, the analysis shows that if f is smaller than a threshold roughly 1/ln(2Δ)^3 (Δ is the maximum node degree) the multiplicative overhead on the convergence time is bounded by

 1 / (1 – p·f·ln(2Δ)^3)

where p is a constant. For realistic values of f this overhead is a small constant; in particular, for f ≤ 1/(e·(2Δ)^e) the overhead becomes constant. However, loss introduces a systematic bias: after convergence the expected estimate lies in