Disaggregated Computing System for Big Data Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current computer architectures face challenges in efficiently processing large data sets, such as those used in natural language processing and machine learning, due to high throughput requirements and energy consumption, leading to expensive and time-inefficient processing.
Innovation Solution
A dynamically composable computing system with disaggregated hardware resources, utilizing interconnects like RDMA and PCIe networks to offload memory and storage from CPUs and IPUs, allowing for scalable storage bandwidth and capacity, and implementing a configurable load store unit and computational register file to optimize memory and storage operations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If general-purpose processors are used for big data processing, then versatility is maintained, but processing throughput and energy efficiency deteriorate
Solution Approach 1:
The system segments processing workloads into different types (compute-intensive, memory-intensive, storage-intensive, I/O-intensive) and assigns them to specialized processing units (CPUs, IPUs, GPUs, FPGAs, storage devices) rather than using general-purpose processors for all tasks. This segmentation enables each component to operate at optimal efficiency for its specific function while maintaining overall system versatility through the resource manager's ability to dynamically allocate different unit types to different workloads.
2Speed
If local memory and storage are tightly coupled to processing units, then access speed is improved, but storage bandwidth and capacity scalability deteriorate
Solution Approach 1:
The system extracts memory and storage resources from their traditional tight coupling with processing units and places them in separate, independently scalable pools. The resource manager orchestrates access between processing units and these disaggregated memory/storage pools, allowing storage capacity and bandwidth to scale independently without being constrained by processing unit limitations, while maintaining fast access through optimized data paths.
3Measurement precision
If multiple calls are made to the register file for complex operations, then computational accuracy is improved, but processing time and energy consumption increase
Solution Approach 1:
The system merges multiple sequential register file access operations into parallel operations by distributing them across multiple processing units simultaneously. The resource manager coordinates these parallel operations to achieve the same computational result with fewer total cycles, reducing processing time and energy consumption while maintaining computational accuracy through coordinated execution.
Data Source
AI summary
Provided are systems and methods for big data processing and related architectures. Various embodiments include a configurable load store unit, a computational register file, and related methods, systems, and devices. Requests to utilize at least one of a memory and a storage can be received at a computing system comprising a local memory and local storage. Systems and methods can determine availability of a remote memory and a remote storage at one or more remote nodes accessible by the computing system, determine a distribution among the local memory, local storage, and one or more remote nodes to fulfill the request, and based on the determination, utilize at least one of: a memory associated with a first set of one or more remote nodes via a first interconnect, and a storage associated with a second set of one or more remote nodes via a second interconnect.