[Paper Review] Implementing $\Diamond P$ with Bounded Messages on a Network of ADD Channels

Saptaparni Kumar, Jennifer L. Welch|arXiv (Cornell University)|Aug 9, 2017

Distributed systems and fault tolerance5 references3 citations

TL;DR

This paper presents a novel algorithm to implement the eventually perfect failure detector ($\Diamond P$) in arbitrary, partitionable networks using only bounded-size messages over unreliable Average Delayed/Dropped (ADD) channels. The method combines bounded heartbeats, dynamic timeout adjustments, and path-based reachability checks to achieve eventual strong accuracy and strong completeness with minimal assumptions on channel reliability and network topology.

ABSTRACT

We present an implementation of the eventually perfect failure detector ($\Diamond P$) from the original hierarchy of the Chandra-Toueg oracles on an arbitrary partitionable network composed of unreliable channels that can lose and reorder messages. Prior implementations of $\Diamond P$ have assumed different partially synchronous models ranging from bounded point-to-point message delay and reliable communication to unbounded message size and known network topologies. We implement $\Diamond P$ under very weak assumptions on an arbitrary, partitionable network composed of Average Delayed/Dropped (ADD) channels to model unreliable communication. Unlike older implementations, our failure detection algorithm uses bounded-sized messages to eventually detect all nodes that are unreachable (crashed or disconnected) from it.

Motivation & Objective

• To design a failure detection algorithm that works under minimal system assumptions, particularly in partitionable networks with unreliable, partially synchronous channels.
• To ensure the failure detector achieves both strong completeness and eventual strong accuracy, even when message delays and losses are unbounded for most messages.
• To maintain bounded message sizes—specifically, messages of size $O((n+1)!)$—while supporting arbitrary network topologies.
• To extend prior work on ADD channels, which previously only applied to fully connected networks, to general, arbitrary topologies.
• To provide a practical and scalable solution for failure detection in real-world distributed systems with unreliable communication.

Proposed method

• Each node periodically sends bounded-size heartbeat messages to its neighbors using privileged messages on ADD channels, which guarantee delivery within a delay bound $d$ and limit unprivileged message bursts.
• Nodes maintain local suspect states and dynamically adjust timeout values based on missed heartbeats, increasing the timeout upon incorrect suspicion to avoid false positives.
• For non-neighboring nodes, the algorithm evaluates reachability by checking all simple paths in the network graph; a node is suspected if no path exists where all intermediate nodes are correctly detected as live.
• The algorithm uses a path enumeration mechanism to track potential routes between nodes, with path information updated only when new suspect states are confirmed.
• The suspect state of a node is set to true if any path to it contains a node already marked as suspect, and this state is never reversed once set.
• The use of bounded data structures—$\texttt{suspect\_local}[\cdot]$ of size $n$ bits and $\texttt{paths}[\cdot]$ of size $O((n+1)!)$—ensures message size remains bounded.

Experimental results

Research questions

• RQ1 Can an eventually perfect failure detector ($\Diamond P$) be implemented in a partitionable network with only bounded message sizes and no assumptions on network topology or channel reliability?
• RQ2 What system assumptions are necessary and sufficient to achieve strong completeness and eventual strong accuracy in failure detection under unbounded message loss and delay?
• RQ3 How can bounded-size messages be used effectively to maintain path-based reachability information in a dynamic, unreliable network?
• RQ4 Can the ADD channel model—previously limited to fully connected networks—be extended to arbitrary, partitionable topologies with bounded message overhead?
• RQ5 Is it possible to achieve perfect failure detection with minimal assumptions on synchrony, channel behavior, and message size?

Key findings

• The algorithm achieves strong completeness: every correct node eventually suspects all nodes that are unreachable (crashed or disconnected) from it.
• The algorithm achieves eventual strong accuracy: every correct node eventually stops suspecting any node that is reachable from it.
• The suspect state of each node is set only once and never reverted, ensuring stability and consistency of failure detection decisions.
• The message size is bounded by $O((n+1)!)$ bits, which is practical for sparsely connected networks where the number of paths is much smaller than the worst-case bound.
• The algorithm is correct under the weakest possible assumptions: arbitrary network topology, unbounded message loss and delay for unprivileged messages, and no requirement for reliable channels.
• The correctness proof relies on inductive reasoning over path distances, showing that nodes at increasing distances from a target eventually gain perfect information about its reachability status.