Speaker
Prof.
Mads Dam
(KTH)
Description
Distributed aggregation is the problem of computing, in a
decentralized and scalable way, global functions of local
values residing at nodes in a network. Interesting
aggregation functions include average, counting, sum,
min/max, voting, medians and quantiles, and thresholds. In
network management applications, our primary domain of
interest, such aggregates can be useful indicators of the
network state. The computation may be subject to constraints
regarding a range of parameters such as accuracy,
responsiveness, robustness to node or link failures, and
security. In the talk I survey recent work at KTH on
distributed aggregation both using tree-based and
gossip-based approaches. A particular difficulty with
gossip-based aggregation is to recover state in case of node
failures. We outline a recently developed solution which is
convergent under the assumption that no two neighbouring
nodes fail within the same round.
