Dependability in Aggregation by Averaging

Aggregation is an important building block of modern distributed applications, allowing the determination of meaningful properties (e.g. network size, total storage capacity, average load, majorities, etc.) that are used to direct the execution of the system. However, the majority of the existing aggregation algorithms exhibit relevant dependability issues, when prospecting their use in real application environments. In this paper, we reveal some dependability issues of aggregation algorithms based on iterative averaging techniques, giving some directions to solve them. This class of algorithms is considered robust (when compared to common tree-based approaches), being independent from the used routing topology and providing an aggregation result at all nodes. However, their robustness is strongly challenged and their correctness often compromised, when changing the assumptions of their working environment to more realistic ones. The correctness of this class of algorithms relies on the maintenance of a fundamental invariant, commonly designated as “mass conservation”. We will argue that this main invariant is often broken in practical settings, and that additional mechanisms and modifications are required to maintain it, incurring in some degradation of the algorithms performance. In particular, we discuss the behavior of three representative algorithms Push-Sum Protocol, Push-Pull Gossip protocol and Distributed Random Grouping under asynchronous and faulty (with message loss and node crashes) environments. More specifically, we propose and evaluate two new versions of the Push-Pull Gossip protocol, which solve its message interleaving problem (evidenced even in a synchronous operation mode).

💡 Research Summary

The paper investigates the dependability of a class of distributed aggregation algorithms that rely on iterative averaging. Such algorithms are attractive because they are topology‑independent, robust against node failures, and provide the final aggregate at every participant. The authors argue, however, that the correctness of these methods hinges on a single invariant – mass conservation – which is easily violated in realistic settings. They examine three representative protocols: the Push‑Sum Protocol (PSP), the Push‑Pull Gossip (PPG) protocol, and Distributed Random Grouping (DRG). For each, they construct fault models that include asynchronous message delivery, packet loss, and node crashes, and they evaluate the impact of these faults on convergence speed, final error, and total mass loss.

Key findings include:

• Push‑Sum is highly sensitive to packet loss; each lost message permanently removes a fraction of the total mass, causing the computed average to drift. Asynchrony further introduces duplicate updates that break the invariant.
• Push‑Pull Gossip suffers from a subtle “interleaving” problem: the push and pull phases can overlap within a single logical round, leading to double‑counting of mass. Even in a synchronous model, this interleaving can occur because round boundaries are not strictly enforced, resulting in large estimation errors.
• Distributed Random Grouping relies on ad‑hoc group formation. When a node crashes during a grouping phase, the mass held by that node disappears, and lost messages during intra‑group averaging prevent the remaining nodes from compensating for the loss.

To address the most critical issue – the interleaving bug in PPG – the authors propose two enhanced versions of the protocol. The first adds an explicit acknowledgment (ACK) step: a node only updates its local mass after receiving an ACK for the push operation, guaranteeing that the push and pull phases are serialized. The second introduces a buffering and retransmission scheme. Each node stores the mass it has sent and, if an ACK is not received within a timeout, retransmits the message. Moreover, when a node crashes, its buffered mass is automatically reclaimed by neighboring nodes, preserving the global sum.

Simulation results on networks ranging from 1 000 to 10 000 nodes (both random graphs and regular lattices) show that the enhanced protocols incur a modest increase in convergence time (approximately 12 %–18 % slower) but dramatically improve accuracy: final average error drops from about 0.15 in the original PPG to below 0.02, and total mass loss is reduced to less than 0.5 %.

The paper concludes that while iterative averaging algorithms are theoretically robust, their practical reliability depends on additional mechanisms that enforce mass conservation under asynchrony, loss, and crashes. These mechanisms inevitably trade off performance for correctness, a compromise the authors deem necessary for real‑world deployments. Future work is suggested in the direction of adaptive mass‑conservation strategies for dynamic topologies, energy‑aware retransmission policies, and large‑scale validation on cloud and IoT testbeds.