Pooled Memory Network Communication in Rack-Scale Architecture
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional data center architectures face limitations in scalability and efficiency due to the rigid allocation of computing resources, leading to challenges in managing thermal issues and power consumption, which affect compute density and performance.
Innovation Solution
The implementation of rack-scale architecture, which disaggregates and pools computing resources such as compute nodes and memory, allowing for dynamic allocation and improving networking throughput by using a pooled memory system that reduces latency and network traffic.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If dedicated computing resources are allocated to each server, then reliability is improved, but device complexity and scalability worsen
Solution Approach 1:
The system segments computing resources by separating compute nodes from memory resources. Each compute node can access a shared pooled memory system rather than requiring dedicated local memory, reducing hardware complexity while maintaining reliable access to necessary resources through the interconnected memory pool.
Solution Approach 2:
The pooled memory system serves multiple compute nodes simultaneously, making the memory infrastructure universal rather than dedicated to single servers. This multi-functional memory pool reduces overall device complexity while ensuring each compute node has reliable access to the shared resource pool.
2Adaptability or versatility
If computing resources are disaggregated and pooled, then scalability is improved, but device complexity increases
Solution Approach 1:
The architecture segments the data center into independent compute nodes and a shared memory pool, allowing each component to scale independently. Compute nodes can be added or removed without affecting the memory infrastructure, enabling scalable deployment while managing complexity through modular design.
Solution Approach 2:
The pooled memory system acts as an intermediary layer between compute nodes and storage resources. This intermediate pooling architecture simplifies scalability by providing a standardized interface that abstracts the complexity of direct resource management, allowing compute nodes to scale without proportionally increasing system complexity.
3Productivity
If rack-scale architecture with pooled memory is implemented, then productivity is improved, but device complexity increases
Solution Approach 1:
The system merges multiple memory resources from different compute nodes into a unified pooled memory system. This consolidation improves productivity by enabling efficient resource utilization and reduced data copying between nodes, while the virtualization layer manages the complexity of the merged resources.
Solution Approach 2:
The pooled memory system provides universal access to all compute nodes within the rack, enabling any node to access any memory resource. This multi-functional architecture improves productivity through flexible resource allocation and reduced latency, while centralized management handles the complexity of universal access coordination.
Data Source
AI summary
Technologies for network communication using pooled memory include a computing rack having a pooled memory enclosure with a pooled memory controller and a compute enclosure having two or more compute nodes. A first compute node determines a destination virtual network interface controller identifier (vNIC ID) associated with a destination network address. The first compute node transmits a send message to the pooled memory controller that includes the destination vNIC ID and a sender physical address of packet data within the pooled memory. The pooled memory controller transmits a receive message to a second compute node associated with the destination vNIC ID. The second compute node transmits a receiver physical address of a receive buffer within the pooled memory to the pooled memory controller. The pooled memory controller copies the packet data from the sender physical address to the receiver physical address. Other embodiments are described and claimed.