On the Feasibility of Wireless Interconnects for High-throughput Data Centers

TL;DR

This paper explores the potential of wireless data center networks using millimeter-wave technology, analyzing throughput scalability under heterogeneous traffic and proposing schemes to improve network performance.

Contribution

It introduces a lower bound for throughput scalability in rate-heterogeneous wireless multi-hop networks and proposes a 2-partitioning scheme to enhance throughput.

Findings

Conventional multi-hopping is inefficient for heterogeneous traffic.

The proposed 2-partitioning scheme improves throughput bounds.

Simulation results validate the theoretical improvements.

Abstract

Data Centers (DCs) are required to be scalable to large data sets so as to accommodate ever increasing demands of resource-limited embedded and mobile devices. Thanks to the availability of recent high data rate millimeter-wave frequency spectrum such as 60GHz and due to the favorable attributes of this technology, wireless DC (WDC) exhibits the potentials of being a promising solution especially for small to medium scale DCs. This paper investigates the problem of throughput scalability of WDCs using the established theory of the asymptotic throughput of wireless multi-hop networks that are primarily proposed for homogeneous traffic conditions. The rate-heterogeneous traffic distribution of a data center however, requires the asymptotic heterogeneous throughput knowledge of a wireless network in order to study the performance and feasibility of WDCs for practical purposes. To answer these questions this paper presents a lower bound for the throughput scalability of a multi-hop rate-heterogeneous network when traffic generation rates of all nodes are similar, except one node. We demonstrate that the throughput scalability of conventional multi-hopping and the spatial reuse of the above bi-rate network is inefficient and henceforth develop a speculative 2-partitioning scheme that improves the network throughput scaling potentials. A better lower bound of the throughput is then obtained. Finally, we obtain the throughput scaling of an i.i.d. rate-heterogeneous network and obtain its lower bound. Again we propose a speculative 2-partitioning scheme to achieve a network with higher throughput in terms of improved lower bound. All of the obtained results have been verified using simulation experiments.