[Paper Review] On Distributed Algorithms for Cost-Efficient Data Center Placement in Cloud Computing

Wuqiong Luo, Wee Peng Tay+2 more|arXiv (Cornell University)|Feb 5, 2018

Caching and Content Delivery17 references3 citations

TL;DR

This paper proposes a fully distributed, linear-complexity algorithm called Distributed Lloyd's Method (DLM) for cost-efficient data center placement in cloud computing. Using iterative Voronoi partitioning and local message-passing, DLM achieves performance comparable to centralized greedy algorithms—despite requiring only local information—on both synthetic and real Internet topologies, with convergence guaranteed on tree networks.

ABSTRACT

The increasing popularity of cloud computing has resulted in a proliferation of data centers. Effective placement of data centers improves network performance and minimizes clients' perceived latency. The problem of determining the optimal placement of data centers in a large network is a classical uncapacitated $k$-median problem. Traditional works have focused on centralized algorithms, which requires knowledge of the overall network topology and information about the customers' service demands. Moreover, centralized algorithms are computationally expensive and do not scale well with the size of the network. We propose a fully distributed algorithm with linear complexity to optimize the locations of data centers. The proposed algorithm utilizes an iterative two-step optimization approach. Specifically, in each iteration, it first partitions the whole network into $k$ regions through a distributed partitioning algorithm; then within each region, it determines the local approximate optimal location through a distributed message-passing algorithm. When the underlying network is a tree topology, we show that the overall cost is monotonically decreasing between successive iterations and the proposed algorithm converges in a finite number of iterations. Extensive simulations on both synthetic and real Internet topologies show that the proposed algorithm achieves performance comparable with that of centralized algorithms that require global information and have higher computational complexity.

Motivation & Objective

• Address the scalability limitations of centralized algorithms for data center placement in large-scale cloud networks.
• Develop a fully distributed algorithm that operates with only local network and demand information, avoiding global topology knowledge.
• Minimize clients’ perceived latency and overall network cost by solving the NP-hard uncapacitated k-median problem in a decentralized manner.
• Ensure convergence and performance guarantees on tree-structured networks while maintaining low computational complexity.

Proposed method

• Formulate the data center placement problem as an uncapacitated k-median problem to minimize total service cost.
• Propose a distributed two-step iterative algorithm: (1) partition the network into k Voronoi regions using a distributed partitioning protocol, and (2) compute local approximate optimal centers via message-passing within each region.
• Leverage the concept of centroidal Voronoi partitions as a necessary condition for optimality in the k-median problem.
• Use a distributed Lloyd’s method variant that updates region boundaries and center locations iteratively using only neighbor-level communication.
• Ensure monotonic cost reduction and finite convergence on tree topologies by exploiting structural properties of the network.
• Implement a scalable, linear-complexity solution that avoids global computation and is suitable for dynamic, large-scale networks.

Experimental results

Research questions

• RQ1 Can a fully distributed algorithm achieve performance close to centralized optimal solutions for data center placement without requiring global network knowledge?
• RQ2 How does the proposed distributed algorithm compare to centralized greedy algorithms in terms of cost and latency across diverse network topologies?
• RQ3 Under what network conditions does the distributed algorithm converge, and what guarantees can be provided on cost reduction?
• RQ4 What is the impact of sequential vs. simultaneous data center selection on load balancing and overall cost?
• RQ5 How does the algorithm perform on real-world Internet topologies such as AS graphs with dynamic demand patterns?

Key findings

• The proposed Distributed Lloyd’s Method (DLM) achieves a total response time ratio of less than 1.22 compared to the centralized greedy algorithm on real Internet topologies (AS graph with 26,475 nodes).
• On grid networks, DLM outperforms the centralized greedy algorithm, achieving a cost ratio below 1.1, attributed to better load balancing from simultaneous center selection.
• On tree topologies, the algorithm guarantees monotonic cost reduction and finite convergence, ensuring stability and correctness.
• DLM performs comparably to centralized algorithms with $O(n^2k)$ complexity, despite having linear time complexity and requiring only local information.
• Simulations on synthetic and real topologies show that DLM maintains high performance even when service demands are estimated dynamically over time.
• The algorithm’s performance is robust across different network types, including small-world and grid networks, with consistent cost ratios below 1.1 in most cases.