Dynamic Demand-Aware Link Scheduling for Reconfigurable Datacenters

TL;DR

The paper tackles the problem of quickly updating demand-aware datacenter topologies implemented via $k$ edge-disjoint matchings, addressing the bottleneck of topology recomputation under changing traffic. It develops a spectrum of algorithms—static, dynamic, batch-dynamic, and hybrid—along with a filtering speedup and a post-processing routine that guarantees a $\frac{1}{3}$-approximation for $k>1$ and can be run standalone. An extensive experimental study on 39 real-world and 176 synthetic traces shows that dynamic and batch-dynamic methods provide significant running-time gains and reduced recourse with only modest losses in solution weight, while post-processing enhances solution quality. The findings offer practical guidance: for small batches and few matchings, batch-apx is strong; for larger $k$ or batch sizes, dynamic or hybrid approaches yield the best balance of speed and weight, suggesting broad applicability to reconfigurable datacenter architectures.

Abstract

Emerging reconfigurable datacenters allow to dynamically adjust the network topology in a demand-aware manner. These datacenters rely on optical switches which can be reconfigured to provide direct connectivity between racks, in the form of edge-disjoint matchings. While state-of-the-art optical switches in principle support microsecond reconfigurations, the demand-aware topology optimization constitutes a bottleneck. This paper proposes a dynamic algorithms approach to improve the performance of reconfigurable datacenter networks, by supporting faster reactions to changes in the traffic demand. This approach leverages the temporal locality of traffic patterns in order to update the interconnecting matchings incrementally, rather than recomputing them from scratch. In particular, we present six (batch-)dynamic algorithms and compare them to static ones. We conduct an extensive empirical evaluation on 176 synthetic and 39 real-world traces, and find that dynamic algorithms can both significantly improve the running time and reduce the number of changes to the configuration, especially in networks with high temporal locality, while retaining matching weight.

Figures (6)

• Figure 1: Two optical switches establishing direct connections between nodes $v_1,\dots,v_8$, as instructed by the network controller, and the demand graph with the corresponding two matchings.
• Figure 2: Updating vs. recomputing a solution from scratch for $k=2$: The two matchings are visualized in dotted green and dashed blue, the line width of an edge corresponds to its weight. Shaded edges have been changed. Center: Initial situation. Two updates arrive, one increases an edge weight , another decreases an edge weight . Left: The result after processing the updates. Only four edges are affected. Right: The potential result after a full recomputation from scratch. The whole network is reconsidered and also unaffected edges may have change, which can be inefficient.
• Figure 3: Weight distributions of the four datasets on a log-log scale.
• Figure 4: Average per-update times (left axis, boxes and $\blacklozenge$) and weight (right axis, $\star$) for $k=8$ on all real-world instances.
• Figure 5: Average per-update times (left axis, boxes and $\blacklozenge$) and weight (right axis, $\star$) for $k=8$ on all split instances.

Theorems & Definitions (10)

• Lemma 1
• Lemma 2
• Lemma 3
• Lemma 4
• Lemma 5
• Lemma 6
• proof
• Corollary 1
• proof
• Lemma 7