Disaggregated Data Center Memory With Near-Far Allocation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Data centers face challenges in efficiently managing and allocating memory resources due to the limitations of traditional memory architectures, which often result in suboptimal performance and increased costs due to the need for dedicated, physically proximate memory solutions.
Innovation Solution
A disaggregated memory architecture is introduced, where compute resources are coupled to both 'near' and 'far' memory via different interfaces, with 'near' memory providing faster access and 'far' memory offering higher capacity, allowing for dynamic allocation and pooling of memory resources across the data center using an optical fabric and dual-mode optical switching infrastructure.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If memory is physically proximate to compute resources (traditional architecture), then access speed is improved, but memory capacity and flexibility are limited
Solution Approach 1:
The patent segments memory resources from compute resources, creating separate memory pools that can be independently managed and allocated. Memory is divided into discrete allocable units that can be dynamically assigned to different compute resources based on demand, rather than being permanently bound to specific processors.
Solution Approach 2:
The patent introduces fabric interfaces and memory management entities as intermediaries between compute resources and memory pools. These intermediaries enable flexible memory allocation while maintaining efficient access paths, bridging the gap between physical separation and logical connectivity.
2Reliability
If dedicated memory is allocated to each compute resource, then access reliability is improved, but resource utilization efficiency deteriorates
Solution Approach 1:
The patent creates universal memory pools that can serve multiple compute resources simultaneously. The same memory resources can be allocated to different compute nodes based on workload requirements, enabling one memory pool to fulfill multiple functions and serve various computational tasks.
Solution Approach 2:
The patent implements dynamic memory allocation where memory assignments can change over time based on computational needs. Memory resources are not statically bound but can be reallocated, expanded, or contracted dynamically through management entities that respond to changing workload demands.
3Quantity of substance
If memory capacity is increased for each compute node, then storage capability is improved, but thermal issues and power consumption worsen
Solution Approach 1:
The patent extracts memory resources from compute nodes and places them in separate, dedicated memory pools. This physical separation removes the thermal burden of high-capacity memory from compute resources, allowing each to be optimized independently for their respective thermal and power requirements.
4Quantity of substance
If high-capacity memory is provided locally, then memory availability is improved, but system complexity and cost increase
Solution Approach 1:
The patent merges individual memory resources from multiple compute nodes into unified memory pools. This consolidation reduces overall system complexity by eliminating redundant memory controllers and management logic at each compute node, while providing equivalent or greater total capacity through shared resources.
Data Source
AI summary
Examples may include sleds for a rack in a data center including physical compute resources and memory for the physical compute resources. The memory can be disaggregated, or organized into near and far memory. A first sled can comprise the physical compute resources and a first set of physical memory resources while a second sled can comprise a second set of physical memory resources. The first set of physical memory resources can be coupled to the physical compute resources via a local interface while the second set of physical memory resources can be coupled to the physical compute resources via a fabric.