Opportunity
Traditional data centers with server-based architecture suffer from “resource stranding” – a server may have fully utilized CPU but under-utilized memory, yet that spare memory cannot be used by other servers. This leads to inefficient resource utilization and insufficient memory channel bandwidth per CPU core. While composable/disaggregated infrastructure solves stranding by creating shared resource pools, workload consolidation remains challenging. Consolidation aims to rearrange workloads from many nodes onto fewer nodes to improve energy efficiency, but migrating workloads incurs network overhead and service downtime. Existing migration techniques (e.g., pre-copy live migration) have non-negligible costs. There is a trade-off between energy savings (fewer active nodes) and migration costs (number of migrations). No existing method optimally balances these two conflicting objectives, especially for disaggregated data centers where workloads can be partially migrated element-by-element.
Technology
This patent presents a method that consolidates workloads by optimizing an objective function that balances two goals: minimizing active nodes (for energy efficiency) and minimizing workload migrations (to reduce cost/downtime). The objective function is: θ × (number of active nodes) + (1-θ) × (number of migrations), where θ is a weight factor from 0 to 1. By varying θ, the method generates a set of Pareto optimal solutions – each solution represents the optimal energy efficiency for a given number of migrations, or vice versa.
A Q-learning based reinforcement learning algorithm solves this optimization. The algorithm defines: state = number of active nodes; action = moving a workload element to another node of the same type; reward = negative change in the objective function. Key innovations include: (1) a dynamic action space that removes actions for already-processed elements and enforces that all parts of a workload stay in the same resource pool; (2) an exception period that consecutively migrates all elements of a workload when moving to a different pool; (3) each epoch starts from the best placement found so far, not from the last state. The algorithm outputs optimal workload placement. The Pareto front can be displayed to users, allowing selection based on preferences (e.g., minimize downtime vs. maximize energy savings). The method works for both server-based and composable/disaggregated architectures.
Advantages
- Balances Two Conflicting Objectives: Optimizes both energy efficiency (fewer active nodes) and migration cost (fewer migrations), providing a true trade-off.
- Pareto Front Visualization: Displays multiple optimal solutions, allowing operators to choose based on their priorities (e.g., minimum downtime or maximum energy saving).
- Handles Partial Migration: Works for disaggregated data centers where workloads can be split into elements and migrated independently, offering more flexibility.
- Superior to Heuristics: Outperforms traditional methods (First-Fit, First-Fit Decreasing, Simulated Annealing) in balancing active nodes and migrations.
- Robust & Scalable: Q-learning approach approximates optimal ILP solutions closely but scales to large problems where ILP becomes NP-hard.
- Dynamic Adjustment: The algorithm adapts action space dynamically and handles resource constraints in real-time.
Applications
- Data Center Energy Management: Reducing the number of active servers/nodes during low-demand periods to save electricity and cooling costs.
- Cloud Provider Optimization: Consolidating virtual machines onto fewer physical hosts to improve utilization and reduce operational expenses.
- Disaggregated Infrastructure Management: Managing workload placement across pooled CPU, memory, GPU, and FPGA resources.
- Live Migration Planning: Determining which workloads to migrate and when, balancing service interruption against energy savings.
- Green Computing Initiatives: Helping data centers meet sustainability goals by minimizing active hardware without excessive performance impact.
