Homepage Inria website

Section: New Results

Distributed Computing


Rendezvous of Anonymous Agents in Trees

In [5] , we study the so-called rendezvous problem in the mobile agent setting in graph environments. In the studied model, two identical (anonymous) mobile agents start from arbitrary nodes of an unknown tree and have to meet at some node. Agents move in synchronous rounds: in each round an agent can either stay at the current node or move to one of its neighbors. We consider deterministic algorithms for this rendezvous task. The main result of our research is a tight trade-off between the optimal time of completing rendezvous and the size of memory of the agents. For agents with k memory bits, we show that optimal rendezvous time is Θ(n+n2/k) in n-node trees. More precisely, if kclogn, for some constant c, we design agents accomplishing rendezvous in arbitrary trees of size n (unknown to the agents) in time O(n+n2/k), starting with arbitrary delay. We also show that no pair of agents can accomplish rendezvous in time o(n+n2/k), even in the class of lines of known length and even with simultaneous start. Finally, we prove that at least logarithmic memory is necessary for rendezvous, even for agents starting simultaneously in a n-node line.

Rendezvous of Distance-Aware Mobile Agents in Unknown Graphs

In [17] , we study the problem of rendezvous of two mobile agents starting at distinct locations in an unknown graph. The agents have distinct labels and walk in synchronous steps. However, the graph is unlabeled and the agents have no means of marking the nodes of the graph and cannot communicate with or see each other until they meet at a node. When the graph is very large, we would like the time to rendezvous to be independent of the graph size and to depend only on the initial distance between the agents and some local parameters such as the degree of the vertices, and the size of the agent's label. It is well known that even for simple graphs of degree Δ, the rendezvous time can be exponential in Δ in the worst case. In this study, we introduce a new version of the rendezvous problem where the agents are equipped with a device that measures its distance to the other agent after every step. We show that these distance-aware agents are able to rendezvous in any unknown graph, in time polynomial in all the local parameters such the degree of the nodes, the initial distance D and the size of the smaller of the two agent labels l=min(l1,l2). Our algorithm has a time complexity of O(Δ(D+logl)) and we show an almost matching lower bound of Ω(Δ(D+logl/logΔ)) on the time complexity of any rendezvous algorithm in our scenario. Further, this lower bound extends existing lower bounds for the general rendezvous problem without distance awareness.

Rendezvous of Heterogeneous Mobile Agents in Edge-Weighted Networks

In [22] , we study the deterministic rendezvous problem in which a pair of heterogeneous agents, differing in the time required to traverse particular edges of the graph, need to meet on an edge or node of the graph. Each of the agents knows the complete topology of the undirected graph and the initial positions of both of the agents. The agent also knows its own traversal times for all of the edges of the graph, but is unaware of the corresponding traversal times for the other agent. In this scenario, we study the time required by the agents to meet, compared to the time T OPT in the offline scenario in which the agents have complete knowledge of each others capabilities. When no additional assumptions are made, we show that rendezvous can be achieved after time O(nT OPT ) in a n-node graph, and that this time is essentially the best possible in some cases. However, the rendezvous time can be reduced to Θ(T OPT ) when the agents are allowed to exchange Θ(n) bits of information at the start of the rendezvous process. We then show that under some natural assumption about the traversal times of edges, the hardness of the heterogeneous rendezvous problem can be substantially decreased, both in terms of time required for rendezvous without communication, and the communication complexity of achieving rendezvous in time Θ(T OPT ).

Rendezvous with Different Speeds

In [32] we introduce the study of the rendezvous problem in the context of agents having different speeds, and present tight and almost tight bounds for this problem, restricted to a ring topology.

Fair Synchronization

A non-blocking implementation of a concurrent object is an implementation that does not prevent concurrent accesses to the internal representation of the object, while guaranteeing the deadlock-freedom progress condition without using locks. Considering a failure free context, G. Taubenfeld has introduced (DISC 2013) a simple modular approach, captured under a new problem called the fair synchronization problem, to transform a non-blocking implementation into a starvation-free implementation satisfying a strong fairness requirement.

This approach is illustrated in [19] with the implementation of a concurrent stack. The spirit of the paper is mainly pedagogical. Its aim is not to introduce new concepts or algorithms, but to show that a powerful, simple, and modular transformation can provide concurrent objects with strong fairness properties.

In [20] , we extend this approach in several directions. It first generalizes the fair synchronization problem to read/write asynchronous systems where any number of processes may crash. Then, it introduces a new failure detector and uses it to solve the fair synchronization problem when processes may crash. This failure detector, denoted 𝑄𝑃 (Quasi Perfect), is very close to, but strictly weaker than, the perfect failure detector. Last but not least, the paper shows that the proposed failure detector 𝑄𝑃 is optimal in the sense that the information on failures it provides to the processes can be extracted from any algorithm solving the fair synchronization problem in the presence of any number of process crash failures.

Wait Free with Advice

In [7] , we motivate and propose a new way of thinking about failure detectors which allows us to define, quite surprisingly, what it means to solve a distributed task wait-free using a failure detector. In our model, the system is composed of computation processes that obtain inputs and are supposed to produce outputs and synchronization processes that are subject to failures and can query a failure detector.

Under the condition that correct synchronization processes take sufficiently many steps, they provide the computation processes with enough advice to solve the given task wait-free: every computation process outputs in a finite number of its own steps, regardless of the behavior of other computation processes.

Every task can thus be characterized by the weakest failure detector that allows for solving it, and we show that every such failure detector captures a form of set agreement. We then obtain a complete classification of tasks, including ones that evaded comprehensible characterization so far, such as renaming or weak symmetry breaking.

Adaptive Register Allocation

In [18] , we give an adaptive algorithm in which processes use multi-writer multi-reader registers to acquire exclusive write access to their own single-writer, multi-reader registers. It is the first such algorithm that uses a number of registers linear in the number of participating processes. Previous adaptive algorithms require at least Θ(n3/2) registers.

Leader Election

Considering the case of homonyms processes (some processes may share the same identifier) on a ring [21] , we give a necessary and sufficient condition on the number of identifiers to enable leader election. We prove that if l is the number of identifiers then message-terminating election is possible if and only if l is greater than the greatest proper divisor of the ring size even if the processes do not know the ring size. If the ring size is known, we propose a process-terminating algorithm exchanging O(nlog(n)) messages that is optimal.

Concurrency and Fault-tolerance

In [15] , we study the connections between self-stabilization and proof-labeling schemes. It follows from the definition of silent self-stabilization, and from the definition of proof-labeling scheme, that if there exists a silent self-stabilizing algorithm using -bit registers for solving a task T, then there exists a proof-labeling scheme for T using registers of at most bits. The first result in this paper is the converse to this statement. We show that if there exists a proof-labeling scheme for a task T, using -bit registers, then there exists a silent self-stabilizing algorithm using registers of at most O(+logn) bits for solving T, where n is the number of processes in the system. Therefore, as far as memory space is concerned, the design of silent self-stabilizing algorithms essentially boils down to the design of compact proof-labeling schemes. The second result in this paper addresses time complexity. We show that, for every task T with k-bits output size in n-node networks, there exists a silent self-stabilizing algorithm solving T in O(n) rounds, using registers of O(n2+kn) bits. Therefore, as far as running time is concerned, every task has a silent self-stabilizing algorithm converging in a linear number of rounds.

In [27] , we study the connections between, on the one hand, asynchrony and concurrency, and, on the other hand, the quality of the expected solution of a distributed algorithm. The state machine approach is a well-known technique for building distributed services requiring high performance and high availability, by replicating servers, and by coordinating client interactions with server replicas using consensus. Indulgent consensus algorithms exist for realistic eventually partially synchronous models, that never violate safety and guarantee liveness once the system becomes synchronous. Unavoidably, these algorithms may never terminate, even when no processor crashes, if the system never becomes synchronous. We propose a mechanism similar to state machine replication, called RC-simulation, that can always make progress, even if the system is never synchronous. Using RC-simulation, the quality of the service will adjust to the current level of asynchrony of the network — degrading when the system is very asynchronous, and improving when the system becomes more synchronous. RC-simulation generalizes the state machine approach in the following sense: when the system is asynchronous, the system behaves as if k+1 threads were running concurrently, where k is a function of the asynchrony. In order to illustrate how the RC-simulation can be used, we describe a long-lived renaming implementation. By reducing the concurrency down to the asynchrony of the system, RC-simulation enables to obtain renaming quality that adapts linearly to the asynchrony.

Quantum Computing

In [1] , we provide illustrative examples of distributed computing problems for which it is possible to design tight lower bounds for quantum algorithms without having to manipulate concepts from quantum mechanics, at all. As a case study, we address the following class of 2-player problems. Alice (resp., Bob) receives a boolean x (resp., y) as input, and must return a boolean a (resp., b) as output. A game between Alice and Bob is defined by a pair (δ,f) of boolean functions. The objective of Alice and Bob playing game (δ,f) is, for every pair (x,y) of inputs, to output values a and b, respectively, satisfying δ(a,b)=f(x,y), in absence of any communication between the two players, but in presence of shared resources. The ability of the two players to solve the game then depends on the type of resources they share. It is known that, for the so-called CHSH game, i.e., for the game ab=xy, the ability for the players to use entangled quantum bits (qubits) helps. We show that, apart from the CHSH game, quantum correlations do not help, in the sense that, for every game not equivalent to the CHSH game, there exists a classical protocol (using shared randomness) whose probability of success is at least as large as the one of any protocol using quantum resources. This result holds for both worst case and average case analysis. It is achieved by considering a model stronger than quantum correlations, the non-signaling model, which subsumes quantum mechanics, but is far easier to handle.

Distributed Decision and Verification


In [12] , we study the power of randomization in the context of locality by analyzing the ability to “boost” the success probability of deciding a distributed language. The main outcome of this analysis is that the distributed computing setting contrasts significantly with the sequential one as far as randomization is concerned. Indeed, we prove that in some cases, the ability to increase the success probability for deciding distributed languages is rather limited.

Model Variants

In a series of papers [14] , [28] , we analyze distributed decision in the context of various models for distributed computing.

In [28] , we carry on the effort to bridging runtime verification with distributed computability, studying necessary conditions for monitoring failure prone asynchronous distributed systems. It has been recently proved that there are correctness properties that require a large number of opinions to be monitored, an opinion being of the form true, false, perhaps, probably true, probably no, etc. The main outcome of this paper is to show that this large number of opinions is not an artifact induced by the existence of artificial constructions. Instead, monitoring an important class of properties, requiring processes to produce at most k different values does require such a large number of opinions. Specifically, our main result is a proof that it is impossible to monitor k-set-agreement in an n-process system with fewer than min{2k,n}+1 opinions. We also provide an algorithm to monitor k-set-agreement with min{2k,n}+1 opinions, showing that the lower bound is tight.

Finally, in [14] , we tackle local distributed testing of graph properties. This framework is well suited to contexts in which data dispersed among the nodes of a network can be collected by some central authority (like in, e.g., sensor networks). In local distributed testing, each node can provide the central authority with just a few information about what it perceives from its neighboring environment, and, based on the collected information, the central authority is aiming at deciding whether or not the network satisfies some property. We analyze in depth the prominent example of checking cycle-freeness, and establish tight bounds on the amount of information to be transferred by each node to the central authority for deciding cycle-freeness. In particular, we show that distributively testing cycle-freeness requires at least logd-1 bits of information per node in graphs with maximum degree d, even for connected graphs. Our proof is based on a novel version of the seminal result by Naor and Stockmeyer (1995) enabling to reduce the study of certain kinds of algorithms to order-invariant algorithms, and on an appropriate use of the known fact that every free group can be linearly ordered.

Voting Systems

In [44] , [38] , we consider a general framework for voting systems with arbitrary types of ballots such as orders of preference, grades, etc. We investigate their manipulability: in what states of the population may a coalition of electors, by casting an insincere ballot, secure a result that is better from their point of view?

We show that, for a large class of voting systems, a simple modification allows to reduce manipulability. This modification is Condorcification: when there is a Condorcet winner, designate her; otherwise, use the original rule.

When electors are independent, for any non-ordinal voting system (i.e. requiring information that is not included in the orders of preferences, for example grades), we prove that there exists an ordinal voting system whose manipulability rate is at most as high and which meets some other desirable properties. Furthermore, this result is also true when voters are not independent but the culture is decomposable, a weaker condition that we define.

Combining both results, we conclude that when searching for a voting system whose manipulability is minimal (in a large class of systems), one can restrict to voting systems that are ordinal and meet the Condorcet criterion.

In [35] , we examine the geometrical properties of the space of expected utilities over a finite set of options, which is commonly used to model the preferences of an agent. We focus on the case where options are assumed to be symmetrical a priori, which is a classical neutrality assumption when studying voting systems. Specifically, we prove that the only Riemannian metric that respects the geometrical properties and the natural symmetries of the utility space is the round metric. Whereas Impartial Culture is widely used in Social Choice literature but limited to ordinal preference, our theoretical result allows to extend it canonically to cardinal preferences.

In [25] , we study the manipulability of voting systems in a real-life experiment: electing the best paper in the conference Algotel 2012. Based on real ballots, we provide a quantitative study of the manipulability, as a function of the voting system used. We show that, even in a situation where all voting systems give the same winner by sincere voting, choosing the voting system is critical, because it has a huge impact on manipulability. In particular, one voting system fare way be better than the others: Instant-Runoff Voting.