Introduction
In large-scale data centers, physical servers (PS) often operate at low utilization, leading to significant waste of power and hardware resources. Virtual Server Consolidation (VSC) is a critical technique that aims to "pack" multiple virtual machines (VMs) onto the minimum number of physical machines possible. However, existing VSC solutions frequently struggle with two major issues: the "Local Optima Trap," where incremental heuristics miss globally efficient configurations, and "Migration Headaches," where the overhead and complex dependencies of moving VMs create execution deadlocks.
To address these challenges, the Virtual Server Consolidation Manager (VSCM) framework is proposed. VSCM utilizes a two-phase approach that combines global optimization with a feedback-driven migration strategy to maximize density while minimizing operational costs.
System Architecture
The VSCM architecture is designed to centralize decision-making while maintaining real-time awareness of the cluster state. As shown in the architecture diagram, the system consists of a Management Server and multiple Computing Servers.
The Control Center within the Management Server houses three core components:
- Info Manager: Collects resource utilization data from sensors on each computing server.
- VMBP Module: Executes the Globally Optimal VM Bin Packing (GOVMBP) algorithm.
- VMM Module: Manages the Feedback-based VM Migration (FVMM) process and interacts with the VM Management Interface to execute migrations.
The Two-Phase Consolidation Framework
Phase 1: Globally Optimal VM Bin Packing (GOVMBP)
Instead of incrementally tweaking the current cluster state, GOVMBP treats the problem as a 2D Bin Packing problem (considering both CPU and Memory). It calculates the absolute minimum number of servers required to host all VMs from a "clean slate" perspective. This results in a target placement configuration known as mid_cluster.
Phase 2: Feedback-based VM Migration (FVMM)
Transitioning from the current state to the globally optimal state might require moving every VM, which is often prohibitively expensive. FVMM introduces a Feedback Mechanism (FM) that filters the proposed migrations. It evaluates whether canceling a specific migration would still allow the cluster to fit within the minimum number of servers. If the migration is deemed non-essential for reaching the server-count goal, it is dropped to reduce overhead.
Managing Migration Constraints
A significant challenge in live consolidation is the dependency between migrations. VSCM categorizes these into sequential and cyclic constraints.
- Sequential Constraints: As seen in Figure (a), VM1 cannot move to PS2 until VM2 vacates it.
- Cyclic Constraints: As seen in Figure (b), a deadlock occurs when VM1 needs PS2 and VM2 needs PS1, but both servers are currently full.
VSCM resolves these using graph theory. By modeling migrations as a directed graph, the system identifies cycles and utilizes "springboard" servers—temporary landing spots with spare capacity—to break deadlocks and ensure the consolidation plan is executable.
Experimental Results
The framework was evaluated using a simulated cluster of 64 physical servers and 100 randomly generated VMs. The results demonstrated significant improvements over the standard First Fit Decreasing (FFD) heuristic:
- Server Reduction: GOVMBP saved an average of 10.22% more physical servers than FFD.
- Migration Efficiency: The Feedback Mechanism reduced the number of migration steps by 67.6% and total migration overhead by 34.7% compared to a non-feedback approach.
- Feasibility: The constraint resolution logic successfully handled all sequential and cyclic dependencies, guaranteeing that every proposed plan could be executed in a live environment.
Conclusion
VSCM provides a robust framework for data center consolidation by decoupling the search for a global optimum from the execution of migrations. By using a clean-slate optimization followed by a feedback-driven filter and graph-based constraint resolution, it achieves superior server density with significantly lower migration costs than traditional heuristics. Future work will focus on validating these results in real-world, heterogeneous Xen-based virtual clusters.
Summary updated