Modern cloud infrastructures rely on sharing network resources among multiple
independent clients, called tenants. Service providers contractually commit to
guaranteeing each tenant a bandwidth allocation defined by a service level
agreement (SLA). However, the dominant transport protocol in these environments
,which is TCP, ignores these contracts: it distributes bandwidth proportionally to
the number of open flows, not to contractual allocations. A tenant that opens
more connections mechanically captures a larger share of network capacity at the
expense of its neighbors. This phenomenon, known as the noisy neighbor problem,
can cause significant performance degradation in production environments.

This thesis addresses the problem at its root, at the network bottleneck,
through three contributions. The first derives per-tenant fluid-flow models for
the two common TCP variants in modern datacenters - NewReno and CUBIC -
directly from their RFC specifications, extending the classical Hollot-Misra
framework to the multi-tenant setting and mathematically characterizing the
coupling between tenant dynamics through the shared queue. The second
contribution presents TAQM (Tenant-Aware Active Queue Management), an
architecture that resolves this coupling by separating each tenant's traffic
into independent queues with weighted fair scheduling, then synthesizes a
per-tenant Linear Quadratic Regulator from the derived models - with
mathematical decoupling analysis. The third contribution introduces a
trust-based adaptive mechanism that monitors each tenant's traffic
predictability online and automatically adjusts the regulator cost matrices,
removing the need for any manual parameter configuration and making the system
fully autonomous. By integrating control theory with confidence-based
adaptation, this work achieves autonomous, tenant-aware network resource
management in multi-tenant datacenters, providing mathematical stability
guarantees and empirical performance improvements which represents a meaningful step toward self-managing datacenter infrastructure.