Section: New Results

Network and Graph Algorithms

Multisource Rumor Spreading with Network Coding

Participants : David Bromberg, Quentin Dufour, Davide Frey.

The last decade has witnessed a rising interest in Gossip protocols in distributed systems. In particular, as soon as there is a need to disseminate events, they become a key functional building block due to their scalability, robustness and fault tolerance under high churn. However, Gossip protocols are known to be bandwidth intensive. A huge amount of algorithms has been studied to limit the number of exchanged messages using different combinations of push/pull approaches. In this work we revisited the state of the art by applying Random Linear Network Coding to further increase performance. In particular, the originality of our approach consists in combining sparse-vector encoding to send our network-coding coefficients and Lamport timestamps to split messages in generations in order to provide efficient gossiping. Our results demonstrate that we are able to drastically reduce bandwidth overhead and dissemination delay compared to the state of the art. We published our results at INFOCOM 2019 [27].

DiagNet: towards a generic, Internet-scale root cause analysis solution

Participants : Loïck Bonniot, François Taïani.

Internet content providers and network operators allocate significant resources to diagnose and troubleshoot problems encountered by end-users, such as service quality of experience degradations. Because the Internet is decentralized, the cause of such problems might lie anywhere between an end-user’s device and the service datacenters. Further, the set of possible problems and causes cannot be known in advance, making it impossible to train a classifier with all combinations of faults, causes and locations. We explored how machine learning can be used for Internet-scale root cause analysis using measurements taken from end-user devices: our solution, DiagNet, is able to build generic models that (i) do not make any assumption on the underlying network topology, (ii) do not require to define the full set of possible causes during training, and (iii) can be quickly adapted to diagnose new services.

DiagNet adapts recent image analysis tactics for system and network metrics, collected from a large and dynamic set of landmark servers. In details, it applies non-overlapping convolutions and global pooling to extract generic information about the analyzed network. This genericness allows to build a general model, that can later be generalized to any Internet service with minimal effort. DiagNet leverages backpropagation attention mechanisms to extend the possible root causes to the set of available metrics, making the model fully extensible. We evaluated DiagNet on geodistributed mockup web services and automated users running in 6 AWS regions, and demonstrated promising root cause analysis capabilities. While this initial work is being reviewed, we are deploying DiagNet for real web services and users to evaluate its performance in a more realistic setup.

Christoph Neumann (InterDigital) actively participated in this work.

Application-aware adaptive partitioning for graph processing systems

Participant : Erwan Le Merrer.

Modern online applications value real-time queries over fresh data models. This is the case for graph-based applications, such as social networking or recommender systems,running on front-end servers in production. A core problem in graph processing systems is the efficient partitioning of the input graph over multiple workers. Recent advances over Bulk Synchronous Parallel processing systems (BSP) enabled computations over partitions on those workers, independently of global synchronization supersteps. A good objective partitioning makes the understanding of the load balancing and communication trade-off mandatory for performance improvement. This work [32] addresses this trade-off through the proposal of an optimization problem, that is to be solved continuously to avoid performance degradation over time. Our simulations show that the design of the software module we propose yields significant performance improvements over the BSP processing model.

This work was done in collaboration with Gilles Trédan (LAAS/CRNS).

How to Spread a Rumor: Call Your Neighbors or Take a Walk?

Participant : George Giakkoupis.

In [28], we study the problem of randomized information dissemination in networks. We compare the now standard push-pull protocol, with agent-based alternatives where information is disseminated by a collection of agents performing independent random walks. In the visit-exchange protocol, both nodes and agents store information, and each time an agent visits a node, the two exchange all the information they have. In the meet-exchange protocol, only the agents store information, and exchange their information with each agent they meet.

We consider the broadcast time of a single piece of information in an $n$-node graph for the above three protocols, assuming a linear number of agents that start from the stationary distribution. We observe that there are graphs on which the agent-based protocols are significantly faster than push-pull, and graphs where the converse is true. We attribute the good performance of agent-based algorithms to their inherently fair bandwidth utilization, and conclude that, in certain settings, agent-based information dissemination, separately or in combination with push-pull, can significantly improve the broadcast time.

The graphs considered above are highly non-regular. Our main technical result is that on any regular graph of at least logarithmic degree, push-pull and visit-exchange have the same asymptotic broadcast time. The proof uses a novel coupling argument which relates the random choices of vertices in push-pull with the random walks in visit-exchange. Further, we show that the broadcast time of meet-exchange is asymptotically at least as large as the other two's on all regular graphs, and strictly larger on some regular graphs.

As far as we know, this is the first systematic and thorough comparison of the running times of these very natural information dissemination protocols.

This work was done in collaboration with Frederik Mallmann-Trenn (MIT) and Hayk Saribekyan (University of Cambridge, UK).