Rack Position Optimization in Large-Scale Heterogeneous Data Centers

TL;DR

This work addresses rack-position optimization in large-scale heterogeneous data centers by jointly minimizing rack movement and fault-tolerance penalties under multi-layer resource constraints. It introduces a two-tier framework that uses a high-level DRL agent with Leader Reward to order rack-type subproblems and a low-level gradient-based heuristic to map racks to positions, enabling scalability to over $10^5$ positions and 100 rack types. The approach outperforms a gradient-based heuristic by about 7% and a MIP solver by over 30% in objective value, achieving 100% problem-solving success within minutes compared to MIP’s hours-long runtimes. The results demonstrate robust, consistent performance across dynamic demands, making the method well-suited for practical deployment in composable, heterogeneous data-center environments.

Abstract

As rapidly growing AI computational demands accelerate the need for new hardware installation and maintenance, this work explores optimal data center resource management by balancing operational efficiency with fault tolerance through strategic rack positioning considering diverse resources and locations. Traditional mixed-integer programming (MIP) approaches often struggle with scalability, while heuristic methods may result in significant sub-optimality. To address these issues, this paper presents a novel two-tier optimization framework using a high-level deep reinforcement learning (DRL) model to guide a low-level gradient-based heuristic for local search. The high-level DRL agent employs Leader Reward for optimal rack type ordering, and the low-level heuristic efficiently maps racks to positions, minimizing movement counts and ensuring fault-tolerant resource distribution. This approach allows scalability to over 100,000 positions and 100 rack types. Our method outperformed the gradient-based heuristic by 7\% on average and the MIP solver by over 30\% in objective value. It achieved a 100\% success rate versus MIP's 97.5\% (within a 20-minute limit), completing in just 2 minutes compared to MIP's 1630 minutes (i.e., almost 4 orders of magnitude improvement). Unlike the MIP solver, which showed performance variability under time constraints and high penalties, our algorithm consistently delivered stable, efficient results - an essential feature for large-scale data center management.

Figures (5)

• Figure 1: Simplified depiction of a rack movement system in heterogeneous data centers. The figure illustrates different rack types, labeled A, B, C, D, E, and F, each consisting of server nodes rich in specific resource types. The color-coded legend at the bottom represents different scopes within the datacenters, each with its resource limits. The figure also highlights placement requests, which indicate the number of racks required to meet resource usage demands. The rack movement system satisfies placement requests and service level objectives (SLOs) by dynamically moving racks or placing new racks within the data centers.
• Figure 2: Our proposed algorithm integrates Leader Reward and adaptation components for rack positioning optimization. It comprises two main modules: an adaptation module (purple box) containing the problem generator and gradient-based heuristic; and a POMO module (green box) featuring an encoder-decoder architecture. The workflow encompasses initialization (red arrows), training (purple arrows), and rollout (green arrows) phases, with data blocks (peach boxes) showing the information passed between functions.
• Figure 3: Comparison of the Proposed Algorithm, gradient Heuristics, and Gurobi across different metrics.
• Figure 4: Top-three reward rankings for Rack Type 1, Rack Type 7, and Rack Type 9 across 80 $\times$ 10 problem instances. Each figure shows the number of instances where the respective rack type, placed in a specific position within the order, achieved a top-three reward ranking (rank_1, rank_2, or rank_3). The x-axis represents the order position of the rack type, while the y-axis shows the number of instances. Rack Type 1 performs best when placed in position 3, Rack Type 7 performs optimally in position 6, and Rack Type 9 achieves the highest rewards when placed in position 8, indicating preferred order positions for each rack type to maximize performance.
• Figure 5: (a) shows the agent’s scores across epochs, with the blue line representing raw scores and the orange line indicating the moving average of the scores over a 25-epoch window. (b) displays the loss values over epochs, where the red line corresponds to raw losses, and the green line represents the moving average of losses. Both metrics illustrate the agent’s learning process, characterized by high variability during early epochs followed by gradual stabilization in later epochs.