Chen Avin

Matthias Bentert, Chen Avin, Stefan Schmid

The performance of distributed applications often critically depends on the interconnecting network or more specifically on its throughput: how fast data can be carried across a network. Over the last years, great progress has been made in understanding demand-oblivious throughput: how fast a given demand matrix describing pairwise communication requirements can be served on a given network. However, surprisingly little is known today about the achievable demand-aware throughput: the throughput on a network topology which can be optimized toward the demand. Such demand-aware networks have recently gained popularity in datacenters and are enabled by emerging reconfigurable optical technologies. In this paper, we are interested in both the achievable demand-aware throughput bounds as well as in the computational complexity of finding a throughput-optimizing network topology. We take a systematic approach and investigate four variants of demand-aware throughput: we analyze, and derive bounds for, two definitions of throughput, the classic throughput usually considered in the literature, and a new generalized definition which we call weak throughput; for each of them, we consider two routing models, a direct one, where demand can only be served on a single hop, and a general one, where multi-hop routing is allowed. Our main result is a separation result which solves an open problem in the literature about the classic throughput definition, showing that demand-aware topologies can outperform demand-oblivious topologies even in the worst case: the demand-aware throughput asymptotically approaches at least 5/8, while it is known that the demand-oblivious throughput is n/(2n-1), which is roughly 1/2. In terms of computational complexity, we show that computing the demand-aware weak throughput is NP-hard, but computing the demand-aware (weak) direct throughput is polynomial-time solvable.

Chen Avin, Zvi Lotker, Shie Mannor et al.

Motivated by dynamic parameter optimization in finite, but large action (configurations) spaces, this work studies the nonstochastic multi-armed bandit (MAB) problem in metric action spaces with oblivious Lipschitz adversaries. We propose ABoB, a hierarchical Adversarial Bandit over Bandits algorithm that can use state-of-the-art existing "flat" algorithms, but additionally clusters similar configurations to exploit local structures and adapt to changing environments. We prove that in the worst-case scenario, such clustering approach cannot hurt too much and ABoB guarantees a standard worst-case regret bound of $O\left(k^{\frac{1}{2}}T^{\frac{1}{2}}\right)$, where $T$ is the number of rounds and $k$ is the number of arms, matching the traditional flat approach. However, under favorable conditions related to the algorithm properties, clusters properties, and certain Lipschitz conditions, the regret bound can be improved to $O\left(k^{\frac{1}{4}}T^{\frac{1}{2}}\right)$. Simulations and experiments on a real storage system demonstrate that ABoB, using standard algorithms like EXP3 and Tsallis-INF, achieves lower regret and faster convergence than the flat method, up to 50% improvement in known previous setups, nonstochastic and stochastic, as well as in our settings.