Current design complexity trends, poor wire scalability, and power limitations argue in favor of highly modular onchip systems. Today's state-of-the-art CMPs already feature up to a hundred discrete cores. With increasing levels of integration, CMPs with hundreds of cores, cache tiles, and specialized accelerators are anticipated in the near future. Meanwhile, server consolidation and cloud computing paradigms have emerged as profit vehicles for exploiting abundant resources of chip-multiprocessors. As multiple, potentially malevolent, users begin to share virtualized resources of a single chip, CMP-level quality-of-service (QOS) support becomes necessary to provide performance isolation, service guarantees, and security. This work takes a topology-aware approach to on-chip QOS. We propose to segregate shared resources, such as memory controllers and accelerators, into dedicated islands (shared regions) of the chip with full hardware QOS support. We rely on a richly connected Multidrop Express Channel (MECS) topology to connect individual nodes to shared regions, foregoing QOS support in much of the substrate and eliminating its respective overheads. We evaluate several topologies for the QOSenabled shared regions, focusing on the interaction between network-on-chip (NOC) and QOS metrics. We explore a new topology called Destination Partitioned Subnets (DPS), which uses a light-weight dedicated network for each destination node. On synthetic workloads, DPS nearly matches or outperforms other topologies with comparable bisection bandwidth in terms of performance, area overhead, energyefficiency, fairness, and preemption resilience.