My PhD thesis proposal:
Routing in Data-Centers: Theoretical Limits and Algorithms [PDF]
My publications:
Minimizing Congestion in Data-Center Networks with Unsplittable Flows. In submission 2025 [PDF]
Summary: This paper seeks to characterize the scale-down factor by which Clos-based data-center networks approximate a macro-switch abstraction connecting all sources to all destinations under the realistic constraint of a single path per flow. The scale-down factor designates the maximum factor by the aggregate demand on any link exceeds the link capacity. On the one hand, the paper proves that the scaling-down factor is at least 3/2 by constructing a particular set of flows such that, for all routings, the aggregate demand routed on any link exceeds the link capacity by a factor of 3/2. On the other hand, the paper shows that the scale-down factor is at most 9/5 by designing a new algorithm that, for all sets of flows, returns a routing for which the aggregate demand routed on any link exceeds the link capacity by a factor of at most 9/5. Therefore, while Clos-based data-center networks cannot approximate a macro-switch abstraction arbitrarily closely, appropriate routing algorithms can ensure this approximation is within reasonable bounds.
Impossibility Results for Data-Center Routing with Congestion Control and Unsplittable Flows. At ACM PODC 2024 [PDF]
Summary: This paper asks whether datacenter networks built from Clos networks can mimic a macro-switch abstraction connecting all sources to all destinations. Many in networking would expect this to be the case, since Clos networks impeccably substituted for macro-switches in the telephone networks of the old days. However, we show that this expectation is wrong. The fact that flows contend among themselves inside the network through congestion control rather than at the entrance of the network through admission control alters the nature of the problem. Among several impossibility results, the paper shows that optimizing routing paths for max-min fair rates inside the data-center may starve some flows in relation to their rates in the macro-switch abstraction.
From Non-Optimal Routing Protocols to Routing on Multiple Optimality Criteria. At IEEE/ACM Transactions on Networking 2022 [PDF]
Summary: This paper extends the theory, the protocols, and the evaluation introduced in the paper below not only to overcome the limitations of BGP in routing optimally in the context of inter-domain routing, but also to open up the ranges of policies choices available there.
Routing on Multiple Optimality Criteria. At ACM SIGCOMM 2020. [PDF]
Summary: This paper addresses the problem of optimal path routing concurrently for multiple optimality criteria, where the goal is to route each flow of data-packets from source to destination via a path that is most appropriate criterion for that flow. What constitute an optimal path depends on the nature of the network and the application sending the flow. Critically, current routing protocols, such as OSPF, BGP, and EIGRP, are not able to find optimal paths, concurrently, for several optimality criteria; in fact, they can only find optimal paths, separately, for criteria with good mathematical properties. Therefore, we present a solution to this problem that involves: (1) developing a new theory, which allows to transform a collection of optimality criteria that do not satisfy good mathematical properties into a dominance criterion that does; the dominance criterion does not particularize a single path, but rather identifies a set of dominant paths including the optimal; (2) designing new routing protocols, which find a set of dominant paths from each source to each destination and allow data-packet forwarding along any of these paths; and (3) evaluating by simulation the protocols developed, thus attesting to their good behavior.
** Best Paper Award **
Featured on APNIC Blog, and Técnico News
My MSc thesis:
Routing on Multiple Optimality Criteria: Theory and Algorithms [PDF]
** IEEE Portugal Outstanding Thesis Award **
Featured on Técnico News