AlveoliNet: An Incrementally Scalable, Cost-Effective, and High-Performance Two-Layer Based Architecture for Data Centers

The escalating count of online services and substantial influx of data triggers the rapid expansion of data center networks (DCNs). DCN should ensure high scalability, low latency, and cost-efficiency, allowing it to adapt swiftly to the dynamic demands of evolving services. The task of constructing a DCN that marries flexibility, cost-effectiveness, low latency, and robust scalability has emerged as a formidable challenge. Driven by this prevailing tendency, this paper proposes a novel server-centric network termed
$AlveoliNet$
.
$AlveoliNet$
is a symmetric network encompassing several clusters formed of two switch layers and two server types: Symmetric and Binary servers. The Symmetric servers have two ports, implying lower cable complexity and construction cost. To perform remarkably with respect to fault tolerance, incremental scalability, and wide bisection width, the Binary servers are endowed with multi-port in accordance with the network level.
$AlveoliNet$
supports incremental incorporation of servers while maintaining its fundamental topological characteristics; its diameter cannot be beyond five hops, irrespective of the DCN size. Additionally, two routing algorithms are designed exclusively for this architecture. The performance evaluations and empirical findings illustrate that
$AlveoliNet$
achieves a harmonious equilibrium between flexibility, remarkable fault-tolerance, incremental scalability, cost-efficiency, and power savings compared to cutting-edge DCN structures. Specifically, the evaluation results demonstrate that for 103,680 nodes operating at 10 GB/s, the construction cost of
$AlveoliNet$
is approximately 28.8%, 75.7%, 74.6%, 43.2%, and 60.5% of the cost of
$Fat$
-
$Tree$
,
$Totoro$
,
$DCell$
,
$BCube$
, and
$LaScaDa$
, respectively. Furthermore, the simulation outcomes reveal
$AlveoliNet$
as a highly promising architecture, surpassing
$Totoro$
,
$DCell$
,
$BCube$
,
$LaScaDa$
, and
$FiConn$
regarding latency by approximately 12%, 14%, 18%, 4%, and 31%, respectively.