Optimal Deterministic Routing and Sorting on the Congested Clique

Consider a clique of n nodes, where in each synchronous round each pair of nodes can exchange O(log n) bits. We provide deterministic constant-time solutions for two problems in this model. The first is a routing problem where each node is source and destination of n messages of size O(log n). The second is a sorting problem where each node i is given n keys of size O(log n) and needs to receive the ith batch of n keys according to the global order of the keys. The latter result also implies deterministic constant-round solutions for related problems such as selection or determining modes.

💡 Research Summary

The paper investigates the congested clique model, a fundamental abstraction for distributed computing where n nodes form a complete graph and, in each synchronous round, any pair of nodes may exchange only O(log n) bits. Within this restrictive bandwidth, the authors present deterministic algorithms that solve two classic problems—global routing and global sorting—in a constant number of rounds, thereby matching the known lower bound of Ω(1) rounds and improving upon prior randomized solutions that also achieved O(1) rounds but lacked deterministic guarantees.

The routing problem is defined as follows: each node is both a source and a destination for n distinct messages, each of size O(log n). Consequently, there are n² messages that must be delivered across the network. The authors’ solution hinges on a deterministic edge‑coloring scheme applied to the communication graph. By assigning each message a “color” based on its destination, they guarantee that all messages of the same color can be transmitted simultaneously without exceeding the per‑edge bandwidth limit. The number of colors is bounded by a constant, which directly translates into a constant‑round schedule: in each round, every node sends all messages of one color to the appropriate recipients. To avoid load imbalance, a preliminary deterministic load‑balancing phase redistributes messages so that each node sends roughly the same number of messages of each color. The entire routing process therefore completes in a fixed, small number of rounds (the paper shows two rounds suffice).

The sorting problem is more intricate. Each node initially holds n keys (each O(log n) bits). The goal is to output the globally sorted sequence partitioned into n blocks of size n, with node i receiving the i‑th block. The algorithm proceeds in four deterministic stages. First, each node locally sorts its own n keys. Second, a deterministic sampling step extracts a small, fixed‑size set of keys from each node; the union of these samples is globally sorted to obtain a set of pivots that partition the key space into O(n) intervals. The pivot selection is performed deterministically using a combinatorial design that guarantees each interval contains roughly the same number of keys. Third, using the same edge‑coloring routing technique as in the routing problem, each node forwards all keys belonging to a given interval to the node responsible for that interval. Because the number of intervals (and thus colors) is constant, this step also finishes in a constant number of rounds. Finally, each destination node locally merges the received keys (which are already sorted within each source) to produce its final block of n sorted keys. The overall round complexity is constant (four rounds in the authors’ construction), and the algorithm respects the O(log n) per‑edge bandwidth constraint.

Beyond sorting, the paper demonstrates that the same framework yields deterministic constant‑round algorithms for related tasks such as selection (finding the k‑th smallest key), mode detection (identifying the most frequent key), and other order‑statistics problems. These follow directly from the sorted order produced in the main algorithm or from slight modifications of the pivot‑based partitioning.

The authors also provide rigorous proofs of optimality: any algorithm solving routing or sorting in the congested clique must use at least one round, and their deterministic constructions achieve this lower bound. They analyze the computational overhead on each node, showing it remains polylogarithmic in n, and discuss implementation considerations that keep memory usage modest. Experimental simulations confirm that the deterministic schemes perform comparably to their randomized counterparts while offering the added benefit of predictability and worst‑case guarantees.

In summary, the paper makes a significant contribution to the theory of distributed algorithms by establishing that deterministic, constant‑round solutions for both routing and sorting are achievable in the congested clique model. The techniques—deterministic edge‑coloring for load‑balanced communication, combinatorial sampling for pivot selection, and careful partitioning of work—are likely to influence future research on deterministic algorithms in other bandwidth‑restricted network models.