Efficient Virtual Data Center Embedding for Disaggregated Data Centers

Description

Data centers are playing an increasingly important role in providing internet services. The nextgeneration data centers are expected to evolve to virtualized and disaggregated architectures in pursuit of high efficiency and flexibility. We aim to provide efficient methods for the virtual data center (VDC) embedding over a disaggregated data center (DDC). VDC provides the capabilities of an actual data center for the client, including computing, storage, and bandwidth. It is based on the combination of server and network virtualization. With server virtualization, computing and storage resources can be abstracted into virtual machines (VMs). Network virtualization can create multiple independent virtual networks over a shared data center network.Traditional data centers are built based on a set of servers, each of which tightly integrates different resources, e.g., CPU and memory. By contrast, in a DDC, different resources are decoupled and assembled into a set of resource pools interconnected with a shared high-speed network. This new architecture has been demonstrated to significantly improve resource flexibility and efficiency.We will provide efficient methods for various models for optimization problems of embedding VDCs over a DDC. The key novelty of this project is that we, for the first time, consider the “disaggregation” nature of the DDC architectures when embedding VDCs. Existing methods for VDC embedding are mainly for server-based data centers (SDCs), where each VM is mapped to one server. By contrast, in this project, we will consider that each VM is mapped to multiple nodes for different resource demands. In addition, we will also consider the latency and bandwidth requirements of inter-resource communications (e.g., CPU-memory communication) within each VM. We will consider a static request scenario, where all VDC requests are known in advance and are made for unlimited time, and dynamic request scenarios, where individual requests arrive and depart randomly. We will use mathematical programming and optimization methodologies, including ILP and heuristic solutions with provable bounds to address these problems. Oneimportant and novel dynamic scenario that we consider is where the costs (OPEX and CAPEX) depend on the system’s state to account for, e.g., compensation to clients for service degradation, and reuse of released resources. For this, we will use stochastic optimization (reinforcement learning) validated by asymptotically optimal decisions based on the restless bandit model.We will also develop a user-friendly software tool with a user interface based on our proposed algorithms that students and future industry partners will use.