Microservices architectures combine the use of fine-grained and independently-scalable services with lightweight communication protocols, such as REST calls over HTTP. Microservices bring flexibility to the development and deployment of application back-ends in the cloud. Applications such as collaborative editing tools require frequent interactions between the front-end running on users' machines and a back-end formed of multiple microservices. User-perceived latencies depend on their connection to microservices, but also on the interaction patterns between these services and their databases. Placing services at the edge of the network, closer to the users, is necessary to reduce user-perceived latencies. It is however difficult to decide on the placement of complete stateful microservices at one specific core or edge location without trading between a latency reduction for some users and a latency increase for the others. We present how to dynamically deploy microservices on a combination of core and edge resources to systematically reduce user-perceived latencies. Our approach enables the split of stateful microservices, and the placement of the resulting splits on appropriate core and edge sites. Koala, a decentralized and resource-driven service discovery middleware, enables REST calls to reach and use the appropriate split, with only minimal changes to a legacy microservices application. Locality awareness using network coordinates further enables to automatically migrate services split and follow the location of the users. We confirm the effectiveness of our approach with a full prototype and an application to ShareLatex, a microservices-based collaborative editing application.

Current cloud computing infrastructures and their management are highly centralized, and therefore they suffer from limitations in terms of network latency, energy consumption, and possible legal restrictions. Decentralizing the Cloud has been recently proposed as an alternative. However, the efficient management of a geographically dispersed platform brings new challenges related to service localization, network utilization and locality-awareness. We here consider a cloud topology composed of many small datacenters geographically dispersed within the backbone network. In this paper, we present the design, development and experimental validation of Koala, a novel overlay network that specifically targets such a geographically distributed cloud platform. The three key characteristics of Koala are laziness, latency-awareness and topology-awareness. By using application traffic, Koala maintains the overlay lazily while it takes locality into account in each routing decision. Although Koala's performance depends on application traffic, through simulation experiments we show that for a uniformly distributed traffic, Koala delivers similar routing complexity and reduced latency compared to a traditional proactive protocol, such as Chord. Additionally, we show that despite its passive maintenance, Koala can appropriately deal with churn by keeping the number of routing failures low, without significantly degrading the routing performance. Finally, we show how such an overlay adapts to a decentralized cloud composed of multiple small datacenters.

Recent increase in demand for next-to-source data processing and low-latency applications has shifted attention from the traditional centralized cloud to more distributed models such as edge computing. In order to fully leverage these models it is necessary to decentralize not only the computing resources but also their management. While a decentralized cloud has various inherent advantages, it also introduces different challenges with respect to coordination and collaboration between resources. A large-scale system with multiple administrative entities requires an overlay network which enables data and service localization based only on a partial view of the network. Numerous existing overlay networks target different properties but they are built in a generic context, without taking into account the specific requirements of a decentralized cloud. In this paper we identify some of these requirements and introduce Koala, a novel overlay network designed specifically to meet them.

Let consider a set of anonymous moving objects to be tracked in a binary sensor network. This article studies the problem of associating deterministically a track revealed by the sensor network with the trajectory of an unique anonymous object, namely the multiple object tracking and identification (MOTI) problem. In our model, the network is represented by a sparse connected graph where each vertex represents a binary sensor and there is an edge between two sensors if an object can pass from one sensed region to another one without activating any other sensor. The difficulty of MOTI lies in the fact that the trajectories of two or more objects can be so close that the corresponding tracks on the sensor network can no longer be distinguished (track merging), thus confusing the deterministic association between an object trajectory and a track. The article presents several results. We first show that MOTI cannot be solved on a general graph of ideal binary sensors even by an omniscient external observer if all the objects can freely move on the graph. Then we describe restrictions that can be imposed a priori either on the graph, on the object movements, or on both, to make the MOTI problem always solvable. In the absence of an omniscient observer, we show how our results can lead to the definition of distributed algorithms that are able to detect when the system is in a state where MOTI becomes unsolvable.

International audience

We consider message-efficient broadcast tolerating Byzantine faults in a multi-hop wireless sensor network. Assuming a grid network where all nodes have a communication range of r, and a single neighborhood contains at most t dishonest and collision-capable (bad) nodes, each with a message budget mf, we investigate the minimum message budget m that each honest (good) node must have in order to achieve reliable broadcast. We consider three cases: (1) mfis known in advance and m is homogeneous among all good nodes; (2) mfis known in advance and m is heterogeneous among good nodes; (3) mfis unknown. For the first two cases, we present possibility results and broadcast protocols that have message costs within twice the lower bound. For the third case, we present a coding scheme that helps verify the integrity of messages at a receiving node without using any cryptographic techniques. This code leads to a reactive local broadcast primitive that has probabilistic reliability guarantees. Combined with a previously proposed scheme, it results in a broadcast protocol for t <; 1/2r(2r + 1) that guarantees reliability with high probability.

We study the problem of shortest-path geographic routing in a static sensor network. Existing algorithms often make routing decisions based on node information in local neighborhoods. However, it is shown by Kuhn et al. that such a design constraint results in a highly undesirable lower bound for routing performance: if a best route has length c, then in the worst case a route produced by any localized algorithm has length Omega(c2), which can be arbitrarily worse than the optimal. We present VIGOR, a visibility-graph-based routing protocol that produces routes of length Theta(c). Our design is based on the construction of a much reduced visibility graph, which guides nodes to find near-optimal paths. The per-node protocol overheads in terms of state information and message transmission depend only on the complexity of the field's large topological features, rather than on the network size. Simulation results show that our protocol dramatically outperforms localized protocols such as GPSR and GOAFR+ in both average and worst cases, with reasonable extra overheads.

Virtual coordinate geographic routing is an appealing geographic routing approach for its ability to work without physical location information. We examine two representative such routing protocols, namely NoGeo and BVR, and show through experiments and theoretical analysis their limitation in adapting to complex field topologies, in particular fields with concave holes. Based on the new insights, we propose a distributed convex partition protocol that divides the field to subareas with convex shapes, using only connectivity information. A new geographic routing protocol, called CONVEX, that builds upon the partitioning protocol is then described. Simulations demonstrate significant performance improvement of the new routing protocol over NoGeo and BVR, in terms of transmission stretch and maintenance overheads.

In this paper, we consider sensor database systems. Sensors are attached to objects and queries on the objects are operated at the sensor network level. Although queries to such a system might be extremely complex, ensuring efficiently basic functionalities such as broadcast or anycast without any central element is not trivial. In this paper, we provide a suite of *-cast (anycast, k-cast, broadcast) functionalities in a fully decentralized manner. More specifically, we present the design and evaluation of SOLIST, a multi-layer structure for sensors, largely inspired from structured peer-to-peer systems providing such functionalities. The main goal of SOLIST is to limit the overall energy consumption. A type is associated to each sensor, and the *-cast functionalities are implemented at a type granularity regardless of the number of types and their distribution within the network. A typical use of such a system is sensor-based stock management. We evaluate SOLIST through simulations and show that SOLIST achieves a reasonable trade-off between performance and energy consumption.

Many different P2P overlay networks providing various functionalities, targeting specific applications, have been proposed in the past five years. It is now reasonable to consider that multiple overlays may be deployed over a large set of nodes so that the most appropriate overlay might be chosen depending on the application. A physical peer may then host several instances of logical peers belonging to different overlay networks. In this paper, we show that the coexistence of a structured P2P overlay and an unstructured one may be leveraged so that, by building one, the other is automatically constructed as well. More specifically, we show that the randomness provided by an unstructured gossip-based overlay can be used to build the routing table of a structured P2P overlay and the randomness in the numerical proximity links in the structured networks provides the random peer sampling required by gossip-based unstructured overlays. In this paper, we show that maintaining the leaf set of Pastry and the proximity links of an unstructured overlay is enough to build the complete overlays. Simulation results comparing our approach with both a Pastry-like system and a gossip-based unstructured overlay show that we significantly reduce the overlay maintenance overhead without sacrificing the performance.

Mutual exclusion is a basic block of distributed synchronization algorithms. One of the challenges in highly distributed environments (like peer-to-peer or Grid configurations) is to provide scalable synchronizations taking into account the hierarchical network topology. This paper proposes hierarchical mutual exclusion algorithms. These algorithms are extensions of the Naimi-Trehel token algorithm, reducing the cost of latency and the number of message exchanges between far hosts. We propose three main extensions : (1) hierarchical proxy-based approach; (2) aggregation of requests; and (3) token preemption by closer hosts. We compared the performance of these algorithms on an emulated Grid testbed. We study the impact of each of the extensions, showing that the combination of them can greatly improve performance of the original algorithm.

Quorum systems are well-known tools that improve the performance and the availability of distributed systems. In this paper we explore their use as a means to achieve low response time for network services that are replicated and accessed over computing grids. To that end, we propose both a quorum construction and a quorum-based state-machine replication algorithm that tolerates crash failures in a partially synchronous model. We show through the evaluation of a real implementation that although simple, this quorum construction and replication algorithm exhibits a response time 20% lower than that of a regular active replication algorithm in appropriate conditions.

International audience

This paper presents DARX, our framework for building applications that provide adaptive fault tolerance. It relies on the fact that multi-agent platforms constitute a very strong basis for decentralized software that is both flexible and scalable, and makes the assumption that the relative importance of each agent varies during the course of the computation. DARX regroups solutions which facilitate the creation of multi-agent applications in a large-scale context. Its most important feature is adaptive replication: replication strategies are applied on a per-agent basis with respect to transient environment characteristics such as the importance of the agent for the computation, the network load or the mean time between failures. Firstly, the interwoven concerns of multi-agent systems and fault-tolerant solutions are put forward. An overview of the DARX architecture follows, as well as an evaluation of its performances. We conclude, after outlining the promising outcomes, by presenting prospective work.

Chandra and Toueg (1996) introduced the concept of unreliable failure detectors, They showed how, by adding these detectors to an asynchronous system, it is possible to solve the Consensus problem. In this paper, we propose a new implementation of a failure detector. This implementation is a variant of the heartbeat failure detector which is adaptable and can support scalable applications. In this implementation we dissociate two aspects: a basic estimation of the expected arrival date to provide a short detection time, and an adaptation of the quality of service according to application needs. The latter is based on two principles: an adaptation layer and a heuristic to adapt the sending period of "I am alive" messages.