Fabric Library for Hardware-Triggered Collective Communication

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing high-performance computing (HPC) systems face challenges in optimizing collective communication operations, leading to increased latency and inefficiency as the number of nodes grows, due to difficulties in managing hardware offload solutions effectively.

Innovation Solution

A hardware-triggered operation implementation paired with a host-based software fabric library is designed to optimize collective communication operations, utilizing a cache for efficient reuse and managing triggered operations to reduce latency and noise from the operating system and application layers.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If hardware offload solutions are used for collective communication operations, then processing throughput is improved, but latency and variability increase as the number of nodes grows

Engineering Contradiction:
Improveprocessing throughputVSAvoidlatency
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

A fabric library acts as an intermediary layer between the application and hardware offload solutions, providing software management of triggered operations. This intermediary coordinates communication patterns across multiple nodes, manages buffer allocation, and optimizes the sequence of operations to reduce latency while maintaining high throughput through efficient hardware utilization

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The system pre-configures triggered operations and buffer structures in advance before collective communication operations execute. By preparing communication patterns, allocating buffers, and setting up triggered operation sequences beforehand, the system reduces runtime latency and variability when actual communication occurs across the network

Inventive Principle:
Principle #10Preliminary action

2Productivity

If more nodes are added to the HPC system, then computational power increases, but communication latency and variability worsen

Engineering Contradiction:
Improvecomputational powerVSAvoidcommunication latency
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The fabric library segments collective communication operations into discrete triggered operations that can be independently managed and optimized. By breaking down large-scale communication patterns into smaller, manageable units with specific trigger conditions, the system can efficiently coordinate activities across numerous nodes without proportionally increasing latency or variability

Inventive Principle:
Principle #1Segmentation

3Productivity

If hardware offload is implemented, then processing efficiency improves, but system complexity increases

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidsystem complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The fabric library serves as a software intermediary that abstracts the complexity of hardware offload management. It provides a standardized interface for applications to access hardware acceleration capabilities without requiring applications to directly manage complex triggered operation configurations, buffer allocations, and synchronization protocols

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The fabric library implements self-service mechanisms by automatically managing triggered operation states, coordinating buffer allocations, and optimizing communication sequences without requiring extensive manual configuration or intervention. This automation reduces the operational complexity of using hardware offload solutions while maintaining high processing efficiency

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP3696674B1Triggered operations for collective communication
Publication Date: 2023.12.27 INTEL CORP
  • EP3696674B1 patent drawingFigure 1
  • EP3696674B1 patent drawingFigure 2
  • EP3696674B1 patent drawingFigure 3

AI summary

Examples include a method of managing storage for triggered operations. The method includes receiving a request to allocate a triggered operation; if there is a free triggered operation, allocating the free triggered operation; if there is no free triggered operation, recovering one or more fired triggered operations, freeing one or more of the recovered triggered operations, and allocating one of the freed triggered operations; configuring the allocated triggered operation; and storing the configured triggered operation in a cache on an input/output (I/O) device for subsequent asynchronous execution of the configured triggered operation.