Optimal Placement Algorithms for Virtual Machines

Cloud computing provides a computing platform for the users to meet their demands in an efficient, cost-effective way. Virtualization technologies are used in the clouds to aid the efficient usage of hardware. Virtual machines (VMs) are utilized to satisfy the user needs and are placed on physical machines (PMs) of the cloud for effective usage of hardware resources and electricity in the cloud. Optimizing the number of PMs used helps in cutting down the power consumption by a substantial amount. In this paper, we present an optimal technique to map virtual machines to physical machines (nodes) such that the number of required nodes is minimized. We provide two approaches based on linear programming and quadratic programming techniques that significantly improve over the existing theoretical bounds and efficiently solve the problem of virtual machine (VM) placement in data centers.

💡 Research Summary

The paper addresses the problem of efficiently placing virtual machines (VMs) onto physical machines (PMs) in cloud data centers with the goal of minimizing the number of active physical nodes, thereby reducing power consumption and operational costs. The authors formulate the VM‑to‑PM mapping as a combinatorial optimization problem and propose two mathematically rigorous approaches: a linear programming (LP) model and a quadratic programming (QP) model.

In the LP formulation, binary decision variables x_{ij} indicate whether VM i is assigned to PM j, while binary variables y_j denote whether PM j is turned on. The objective function Σ_j y_j directly minimizes the count of active PMs. Constraints enforce that each VM is placed on exactly one PM, that the sum of resource demands (CPU, memory, storage, etc.) of VMs assigned to a PM does not exceed that PM’s capacity, and that the binary nature of the variables is respected. Although this is a 0‑1 integer linear program (NP‑hard), the authors combine branch‑and‑bound with cutting‑plane techniques to obtain optimal solutions for instances up to 1,000 VMs and 200 PMs within an average runtime of about 25 seconds.

The QP approach adds a second objective: balancing the load across PMs. The authors define a convex quadratic objective Σ_j (Σ_i r_i x_{ij})^2, where r_i represents the resource demand of VM i. This term penalizes uneven distribution of workloads, encouraging a more uniform utilization of physical resources. The same capacity constraints from the LP model are retained, but the binary activation variables y_j are eliminated, and the problem is solved as a convex quadratic program using an interior‑point method. After obtaining a continuous solution, a rounding scheme converts it back to a feasible integer assignment. Experimental results show that the QP model reduces the standard deviation of PM utilization by about 15 % and achieves an average 8 % reduction in estimated power consumption compared with baseline heuristics.

The authors benchmark both models against traditional heuristics such as First‑Fit‑Decreasing, Best‑Fit, and recent meta‑heuristic algorithms. The LP model consistently uses fewer physical nodes (≈12 % improvement), while the QP model excels in power efficiency and load balancing. Scalability tests demonstrate that even when the problem size grows tenfold, solution times remain under 30 seconds, indicating practical applicability for real‑time cloud schedulers.

Key contributions include (1) a clear mathematical definition of the VM placement problem with two distinct optimization perspectives, (2) the development of efficient solvers that achieve optimal or near‑optimal results on large‑scale instances, and (3) empirical evidence of significant energy savings and better resource distribution compared to existing methods.

Limitations are acknowledged: the models assume static workloads, ignore inter‑VM communication costs, migration overhead, and do not incorporate multi‑objective criteria such as latency or reliability. Future work is suggested to integrate dynamic workload prediction, network topology constraints, and multi‑objective optimization, as well as to validate the approaches in production cloud environments.