Opportunity
Traditional data centers (DCs) face significant inefficiencies due to resource stranding in server-based architectures, where hardware components like CPUs and memory are tightly coupled within monolithic servers. This coupling leads to underutilization of resources, as unused components in one server cannot be reallocated to other tasks. Additionally, hardware failures in such architectures often result in cascading failures, disrupting multiple services. The rapid growth of cloud computing, big data, and edge computing has exacerbated these issues, placing immense pressure on DCs to improve resource efficiency and reliability. Disaggregated Data Centers (DDCs) have emerged as a promising solution by decoupling hardware resources into pools interconnected via networks. However, DDCs introduce new challenges, such as managing inter-resource communication latency, bandwidth constraints, and the reliability of shared network components. This patent addresses these challenges by proposing a novel resource scheduling framework that optimizes both resource utilization and service reliability in DDCs.
Technology
The patent introduces a server and resource scheduling method designed for DDCs with a composable/disaggregated infrastructure. The server comprises multiple processing modules interconnected via communication links, forming a network with a DDC architecture. A DCC hardware monitor detects real-time hardware information, including topology, load, and failure/repair status. A task scheduler module analyzes resource allocation requests and hardware data to allocate tasks dynamically. Key innovations include:
- Static Scenario Handling: Uses Mixed-Integer Linear Programming (MILP) to batch-process requests, maximizing acceptance ratios and per-request reliability by optimizing resource allocation across disaggregated components.
- Dynamic Scenario Handling: Implements heuristic algorithms to schedule incoming requests in real-time, accepting requests if resources are available or blocking them otherwise. A migration-based restoration mechanism reallocates resources during hardware failures.
- Multi-Objective Optimization: Balances resource efficiency (e.g., minimizing stranding) and reliability (e.g., minimizing service interruptions) by considering network latency, bandwidth, and hardware failure probabilities.
Advantages
- Improved Resource Utilization: Decouples hardware components, enabling flexible allocation and reducing stranding.
- Enhanced Reliability: Proactively monitors hardware states and migrates tasks during failures to minimize downtime.
- Scalability: Supports rack-, pod-, and DC-scale disaggregation, adapting to varying latency/bandwidth requirements.
- Cost Efficiency: Reduces hardware upgrade costs by enabling component-level upgrades instead of server replacements.
Application
- Cloud and Edge Computing: Optimizes resource allocation for virtual machines, containers, and distributed workloads.
- High-Performance Computing (HPC): Enhances reliability for latency-sensitive tasks like AI/ML training.
- Disaster Recovery: Mitigates service disruptions via dynamic migration during hardware failures.
- 5G and IoT Networks: Manages resource demands for low-latency applications in fog/edge environments.
For more information, please visit: https://www.ee.cityu.edu.hk/~zukerman/
