GPU Cluster Network Message Aggregation and Compression

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

In computing systems, particularly those with GPUs, the reliance on CPUs for network communication leads to delays and inefficiencies due to the lack of support for I/O abstractions, resulting in high overhead when transferring numerous small network messages across the network.

Innovation Solution

A processing node with a heterogeneous multi-core architecture combines and compresses network messages using control logic to reduce the number and size of messages, leveraging operation combiners, data compressors, and message aggregators to generate new messages with shared data stored only once, thereby reducing network traffic and latency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If parallel processors generate independent network messages for each work-item, then each processor can communicate directly, but the number of small messages increases significantly causing high overhead

Engineering Contradiction:
Improvedirect network communication capabilityVSAvoidnetwork message overhead
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent merges multiple small network messages generated by parallel processors into fewer aggregated messages. The network interface controller combines messages from multiple work-items that share the same destination, reducing the total number of messages transmitted across the network and eliminating redundant overhead.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The network interface controller performs multiple functions: it acts as both a communication interface and a message aggregation point. By consolidating the message aggregation function at the network interface, the system eliminates the need for CPU intervention in message combining while maintaining direct processor-to-network communication capability.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Adaptability or versatility

If CPU coordinates network communications for parallel processors, then network communication is supported, but processor availability decreases due to CPU occupation

Engineering Contradiction:
Improvenetwork communication supportVSAvoidprocessor availability
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The system enables self-service network communication by empowering the network interface controller to autonomously aggregate and manage messages from parallel processors without CPU intervention. The processors generate messages independently, and the network interface handles aggregation and transmission, allowing processors to remain available for computation while network communication proceeds independently.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The network interface controller serves as an intermediary between parallel processors and the network. It receives messages from multiple processors, aggregates them by destination, and transmits consolidated messages, eliminating the need for CPU mediation and allowing processors to focus on computation while network communication is handled autonomously.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Reliability

If multiple identical messages are sent across the network, then all destinations receive the data, but network traffic increases unnecessarily

Engineering Contradiction:
Improvemessage delivery to multiple destinationsVSAvoidnetwork traffic
Core Design Contradiction:
ReliabilityVSLoss of energy

Solution Approach 1:

The patent combines multiple identical or similar messages destined for different locations into a single aggregated message. The network interface controller identifies messages with the same source and destination node, consolidates them, and transmits one combined message, thereby delivering data to multiple destinations efficiently while reducing network traffic and energy consumption.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentUS10320695B2Message aggregation, combining and compression for efficient data communications in GPU-based clusters
Publication Date: 2019.06.11 ADVANCED MICRO DEVICES INC
  • US10320695B2 patent drawing
  • US10320695B2 patent drawing
  • US10320695B2 patent drawing

AI summary

A system and method for efficient management of network traffic management of highly data parallel computing. A processing node includes one or more processors capable of generating network messages. A network interface is used to receive and send network messages across a network. The processing node reduces at least one of a number or a storage size of the original network messages into one or more new network messages. The new network messages are sent to the network interface to send across the network.