Section: New Results

Distributed Computing

On the Limits of Noise in Distributed Computing

Biological systems can share and collectively process information to yield emergent effects, despite inherent noise in communication. While man-made systems often employ intricate structural solutions to overcome noise, the structure of many biological systems is more amorphous. It is not well understood how communication noise may affect the computational repertoire of such groups. To approach this question we consider in [9], [15] the basic collective task of rumor spreading, in which information from few knowledgeable sources must reliably flow into the rest of the population. We study the effect of communication noise on the ability of groups that lack stable structures to efficiently solve this task. We present an impossibility result which strongly restricts reliable rumor spreading in such groups. Namely, we prove that, in the presence of even moderate levels of noise that affect all facets of the communication, no scheme can significantly outperform the trivial one in which agents have to wait until directly interacting with the sources—a process which requires linear time in the population size. Our results imply that in order to achieve efficient rumor spread a system must exhibit either some degree of structural stability or, alternatively, some facet of the communication which is immune to noise. We then corroborate this claim by providing new analyses of experimental data regarding recruitment in Cataglyphis niger desert ants. Finally, in light of our theoretical results, we discuss strategies to overcome noise in other biological systems.

Minimizing message size in stochastic communication patterns: fast self-stabilizing protocols with 3 bits

In [8], we consider the basic PULL model of communication, in which in each round, each agent extracts information from few randomly chosen agents. We seek to identify the smallest amount of information revealed in each interaction (message size) that nevertheless allows for efficient and robust computations of fundamental information dissemination tasks. We focus on the Majority Bit Dissemination problem that considers a population of $n$ agents, with a designated subset of source agents. Each source agent holds an input bit and each agent holds an output bit. The goal is to let all agents converge their output bits on the most frequent input bit of the sources (the majority bit). Note that the particular case of a single source agent corresponds to the classical problem of Broadcast (also termed Rumor Spreading). We concentrate on the severe fault-tolerant context of self-stabilization, in which a correct configuration must be reached eventually, despite all agents starting the execution with arbitrary initial states. In particular, the specification of who is a source and what is its initial input bit may be set by an adversary.

We first design a general compiler which can essentially transform any self-stabilizing algorithm with a certain property (called “the the bitwise-independence property”) that uses $\ell$-bits messages to one that uses only $log\ell$-bits messages, while paying only a small penalty in the running time. By applying this compiler recursively we then obtain a self-stabilizing Clock Synchronization protocol, in which agents synchronize their clocks modulo some given integer $T$, within $\tilde{O}(lognlogT)$ rounds w.h.p., and using messages that contain 3 bits only. We then employ the new Clock Synchronization tool to obtain a self-stabilizing Majority Bit Dissemination protocol which converges in $\tilde{O}(logn)$ time, w.h.p., on every initial configuration, provided that the ratio of sources supporting the minority opinion is bounded away from half. Moreover, this protocol also uses only 3 bits per interaction.

Intense Competition can Drive Selfish Explorers to Optimize Coverage

In [30], we consider a game-theoretic setting in which selfish individuals compete over resources of varying quality. The motivating example is a group of animals that disperse over patches of food of different abundances. In such scenarios, individuals are biased towards selecting the higher quality patches, while, at the same time, aiming to avoid costly collisions or overlaps. Our goal is to investigate the impact of collision costs on the parallel coverage of resources by the whole group.

Consider $M$ sites, where a site $x$ has value $f\left(x\right)$. We think of $f\left(x\right)$ as the reward associated with site $x$, and assume that if a single individual visits $x$ exclusively, it receives this exact reward. Typically, we assume that if $\ell >1$ individuals visit $x$ then each receives at most $f\left(x\right)/\ell$. In particular, when competition costs are high, each individual might receive an amount strictly less than $f\left(x\right)/\ell$, which could even be negative. Conversely, modeling cooperation at a site, we also consider cases where each one gets more than $f\left(x\right)/\ell$. There are $k$ identical players that compete over the rewards. They independently act in parallel, in a one-shot scenario, each specifying a single site to visit, without knowing which sites are explored by others. The group performance is evaluated by the expected coverage, defined as the sum of $f\left(x\right)$ over all sites that are explored by at least one player. Since we assume that players cannot coordinate before choosing their site we focus on symmetric strategies.

The main takeaway message of this paper is that the optimal symmetric coverage is expected to emerge when collision costs are relatively high, so that the following “Judgment of Solomon” type of rule holds: If a single player explores a site $x$ then it gains its full reward $f\left(x\right)$, but if several players explore it, then neither one receives any reward. Under this policy, it turns out that there exists a unique symmetric Nash Equilibrium strategy, which is, in fact, evolutionary stable. Moreover, this strategy yields the best possible coverage among all symmetric strategies. Viewing the coverage measure as the social welfare, this policy thus enjoys a (Symmetric) Price of Anarchy of precisely 1, whereas, in fact, any other congestion policy has a price strictly greater than 1.

Our model falls within the scope of mechanism design, and more precisely in the area of incentivizing exploration. It finds relevance in evolutionary ecology, and further connects to studies on Bayesian parallel search algorithms.

Universal Protocols for Information Dissemination Using Emergent Signals

In [23], we consider a population of $n$ agents which communicate with each other in a decentralized manner, through random pairwise interactions. One or more agents in the population may act as authoritative sources of information, and the objective of the remaining agents is to obtain information from or about these source agents. We study two basic tasks: broadcasting, in which the agents are to learn the bit-state of an authoritative source which is present in the population, and source detection, in which the agents are required to decide if at least one source agent is present in the population or not.

We focus on designing protocols which meet two natural conditions: (1) universality, i.e., independence of population size, and (2) rapid convergence to a correct global state after a reconfiguration, such as a change in the state of a source agent. Our main positive result is to show that both of these constraints can be met. For both the broadcasting problem and the source detection problem, we obtain solutions with a convergence time of $O\left({log}^{2}n\right)$ rounds, w.h.p., from any starting configuration. The solution to broadcasting is exact, which means that all agents reach the state broadcast by the source, while the solution to source detection admits one-sided error on a $\epsilon$-fraction of the population (which is unavoidable for this problem). Both protocols are easy to implement in practice and have a compact formulation.

Our protocols exploit the properties of self-organizing oscillatory dynamics. On the hardness side, our main structural insight is to prove that any protocol which meets the constraints of universality and of rapid convergence after reconfiguration must display a form of non-stationary behavior (of which oscillatory dynamics are an example). We also observe that the periodicity of the oscillatory behavior of the protocol, when present, must necessarily depend on the number ${}^{\#}X$ of source agents present in the population. For instance, our protocols inherently rely on the emergence of a signal passing through the population, whose period is $\Theta (log\frac{n}{{}^{\#}X})$ rounds for most starting configurations. The design of clocks with tunable frequency may be of independent interest, notably in modeling biological networks.

Ergodic Effects in Token Circulation

In [25], we consider a dynamical process in a network which distributes all particles (tokens) located at a node among its neighbors, in a round-robin manner.

We show that in the recurrent state of this dynamics (i.e., disregarding a polynomially long initialization phase of the system), the number of particles located on a given edge, averaged over an interval of time, is tightly concentrated around the average particle density in the system. Formally, for a system of $k$ particles in a graph of $m$ edges, during any interval of length $T$, this time-averaged value is $k/m\pm \tilde{O}(1/T)$, whenever $gcd(m,k)=\tilde{O}\left(1\right)$ (and so, e.g., whenever $m$ is a prime number). To achieve these bounds, we link the behavior of the studied dynamics to ergodic properties of traversals based on Eulerian circuits on a symmetric directed graph. These results are proved through sum set methods and are likely to be of independent interest.

As a corollary, we also obtain bounds on the idleness of the studied dynamics, i.e., on the longest possible time between two consecutive appearances of a token on an edge, taken over all edges. Designing trajectories for $k$ tokens in a way which minimizes idleness is fundamental to the study of the patrolling problem in networks. Our results immediately imply a bound of $\tilde{O}(m/k)$ on the idleness of the studied process, showing that it is a distributed $\tilde{O}\left(1\right)$-competitive solution to the patrolling task, for all of the covered cases. Our work also provides some further insights that may be interesting in load-balancing applications.

Improved Analysis of Deterministic Load-Balancing Schemes

In [7], we consider the problem of deterministic load balancing of tokens in the discrete model. A set of $n$ processors is connected into a $d$-regular undirected network. In every time step, each processor exchanges some of its tokens with each of its neighbors in the network. The goal is to minimize the discrepancy between the number of tokens on the most-loaded and the least-loaded processor as quickly as possible.

Rabani et al. (1998) present a general technique for the analysis of a wide class of discrete load balancing algorithms. Their approach is to characterize the deviation between the actual loads of a discrete balancing algorithm with the distribution generated by a related Markov chain. The Markov chain can also be regarded as the underlying model of a continuous diffusion algorithm. Rabani et al. showed that after time $T=O(log(Kn)/\mu )$, any algorithm of their class achieves a discrepancy of $O(dlogn/\mu )$, where $\mu$ is the spectral gap of the transition matrix of the graph, and $K$ is the initial load discrepancy in the system.

In this work we identify some natural additional conditions on deterministic balancing algorithms, resulting in a class of algorithms reaching a smaller discrepancy. This class contains well-known algorithms, eg., the Rotor-Router. Specifically, we introduce the notion of cumulatively fair load-balancing algorithms where in any interval of consecutive time steps, the total number of tokens sent out over an edge by a node is the same (up to constants) for all adjacent edges. We prove that algorithms which are cumulatively fair and where every node retains a sufficient part of its load in each step, achieve a discrepancy of $O(min\{d\sqrt{logn/\mu },d\sqrt{n}\})$ in time $O\left(T\right)$. We also show that in general neither of these assumptions may be omitted without increasing discrepancy. We then show by a combinatorial potential reduction argument that any cumulatively fair scheme satisfying some additional assumptions achieves a discrepancy of $O\left(d\right)$ almost as quickly as the continuous diffusion process. This positive result applies to some of the simplest and most natural discrete load balancing schemes.

The assignment problem

In the allocation problem, asynchronous processors must partition a set of items so that each processor leave knowing all items exclusively allocated to it. In [21], we introduce a new variant of the allocation problem called the assignment problem, in which processors might leave having only partial knowledge of their assigned items. The missing items in a processor's assignment must eventually be announced by other processors.

While allocation has consensus power 2, we show that the assignment problem is solvable read-write wait-free when $k$ processors compete for at least $2k-1$ items. Moreover, we propose a long-lived read-write wait-free assignment algorithm which is fair, allocating no more than 2 items per processor, and in which a slow processor may delay the assignment of at most $n$ items, where n is the number of processors.

The assignment problem and its read-write solution may be of practical interest for implementing resource allocators and work queues, which are pervasive concurrent programming patterns, as well as stream-processing systems.

A Characterization of $t$-Resilient Colorless Task Anonymous Solvability

One of the central questions in distributed computability is characterizing the tasks that are solvable in a given system model. In the anonymous case, where processes have no identifiers and communicate through multi-writer/multi-reader registers, there is a recent topological characterization (Yanagisawa 2017) of the colorless tasks that are solvable when any number of asynchronous processes may crash. In [22], we consider the case where at most $t$ asynchronous processes may crash, where $1\le t<n$. We prove that a colorless task is t-resilient solvable anonymously if and only if it is t-resilient solvable non-anonymously. We obtain our results through various reductions and simulations that explore how to extend techniques for non-anonymous computation to anonymous one.

Implementing Snapshot Objects on Top of Crash-Prone Asynchronous Message-Passing Systems

In asynchronous crash-prone read/write shared-memory systems there is the notion of a snapshot object, which simulates the behavior of an array of single-writer/multi-reader (SWMR) shared registers that can be read atomically. Processes in the system can access the object invoking (any number of times) two operations, denoted write() and snapshot(). A process invokes write() to update the value of its register in the array. When it invokes snapshot(), the process obtains the values of all registers, as if it read them simultaneously. It is known that a snapshot object can be implemented on top of SWMR registers, tolerating any number of process failures. Snapshot objects provide a level of abstraction higher than individual SWMR registers, and they simplify the design of applications. Building a snapshot object on an asynchronous crash-prone message-passing system has similar benefits. The object can be implemented by using the known simulations of a SWMR shared memory on top of an asynchronous message-passing system (if less than half the processes can crash), and then build a snapshot object on top of the simulated SWMR memory. [10] presents an algorithm that implements a snapshot object directly on top of the message-passing system, without building an intermediate layer of a SWMR shared memory. To the authors knowledge, the proposed algorithm is the first providing such a direct construction. The algorithm is more efficient than the indirect solution, yet relatively simple.

Distributed decision

We have carried out our study of distributed decision, either for its potential application to the design of fault-tolerant distributed algorithm, or for the purpose of designing a complexity/computanility theory for distributed network computing.

In the framework of distributed network computing, it is known that not all Turing-decidable predicates on labeled networks can be decided locally whenever the computing entities are Turing machines (TM), and this holds even if nodes are running non-deterministic Turing machines (NTM). In contrast, we show in [6] that every Turing-decidable predicate on labeled networks can be decided locally if nodes are running alternating Turing machines (ATM). More specifically, we show that, for every such predicate, there is a local algorithm for ATMs, with at most two alternations, that decides whether the actual labeled network satisfies that predicate. To this aim, we define a hierarchy of classes of decision tasks, where the lowest level contains tasks solvable with TMs, the first level those solvable with NTMs, and the level $k>1$ contains those tasks solvable with ATMs with $k-1$ alternations. We characterize the entire hierarchy, and show that it collapses in the second level. In addition, we show separation results between the classes of network predicates that are locally decidable with TMs, NTMs, and ATMs, and we establish the existence of completeness results for each of these classes, using novel notions of local reduction. We complete these results by a study of the local decision hierarchy when certificates are bounded to be of logarithmic size.

Distributed proofs are mechanisms enabling the nodes of a network to collectively and efficiently check the correctness of Boolean predicates on the structure of the network (e.g. having a specific diameter), or on data structures distributed over the nodes (e.g. a spanning tree). In [24], we consider well known mechanisms consisting of two components: a prover that assigns a certificate to each node, and a distributed algorithm called verifier that is in charge of verifying the distributed proof formed by the collection of all certificates. We show that many network predicates have distributed proofs offering a high level of redundancy, explicitly or implicitly. We use this remarkable property of distributed proofs to establish perfect tradeoffs between the size of the certificate stored at every node, and the number of rounds of the verification protocol.

The role of unique node identifiers in network computing is well understood as far as symmetry breaking is concerned. However, the unique identifiers also leak information about the computing environment—in particular, they provide some nodes with information related to the size of the network. It was recently proved that in the context of local decision, there are some decision problems that cannot be solved without unique identifiers, but unique identifiers leak a sufficient amount of information such that the problem becomes solvable (PODC 2013). In [11], we give a complete picture of what is the minimal amount of information that we need to leak from the environment to the nodes in order to solve local decision problems. Our key results are related to scalar oracles $f$ that, for any given $n$, provide a multiset $f\left(n\right)$ of $n$ labels; then the adversary assigns the labels to the $n$ nodes in the network. This is a direct generalisation of the usual assumption of unique node identifiers. We give a complete characterisation of the weakest oracle that leaks at least as much information as the unique identifiers. Our main result is the following dichotomy: we classify scalar oracles as large and small, depending on their asymptotic behaviour, and show that (1) any large oracle is at least as powerful as the unique identifiers in the context of local decision problems, while (2) for any small oracle there are local decision problems that still benefit from unique identifiers.