Address Template Identifier for HPC Memory Access
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High Performance Computing (HPC) communication patterns involving small messages reduce system fabric bandwidth due to large header overheads, despite small data payloads, leading to inefficient network operations.
Innovation Solution
Implementing a shared library of address templates across network nodes to replace multiple small messages with an identifier at the sender side, allowing the receiver to reconstruct memory operations using the address template, thereby reducing message traffic and improving bandwidth efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If small messages are used to target multiple destinations on the system fabric, then communication flexibility and addressing capability are improved, but message bandwidth efficiency deteriorates due to large header overhead
Solution Approach 1:
The patent creates a simplified copy or reference (address template identifier) that represents a complete set of memory operation parameters. Instead of transmitting full address information in each small message, the system transmits a compact identifier that references a pre-defined address template containing the complete addressing information. This copying approach maintains communication flexibility while dramatically reducing message header overhead and improving bandwidth efficiency.
Solution Approach 2:
The system performs preliminary action by pre-defining address templates that contain complete sets of memory operation parameters before actual communication occurs. These templates are registered and stored in the system, allowing subsequent small messages to reference them by identifier rather than containing full addressing information. This preliminary preparation enables efficient communication while maintaining the ability to target multiple destinations.
2Measurement precision
If multiple small messages are transmitted to implement memory operations, then precise memory addressing capability is improved, but network bandwidth utilization deteriorates
Solution Approach 1:
The patent uses address templates as pre-defined patterns that can be referenced by compact identifiers. When memory operations need to be performed, instead of transmitting multiple detailed messages with full addressing information, the system transmits a single message containing an address template identifier. The receiving node then uses this identifier to retrieve the complete addressing pattern from the template, achieving precise memory addressing while minimizing network traffic and maximizing bandwidth utilization.
3Reliability
If full address information is included in each message, then message completeness and operation accuracy are improved, but message size and header overhead increase
Solution Approach 1:
The patent implements a reference mechanism where complete address information is stored in pre-defined address templates, and messages contain only compact identifiers referencing these templates. This copying approach ensures operation accuracy is maintained because the full addressing information is preserved in the templates, while message size is dramatically reduced by transmitting only the compact identifier rather than complete address data in each message.
Solution Approach 2:
The address templates serve as universal patterns that can be referenced by multiple messages. A single address template can be used to define addressing patterns for multiple different memory operations, allowing the system to maintain complete addressing information in a reusable format while keeping individual messages compact. This multi-functionality enables both operation accuracy and reduced message size.
Data Source
AI summary
A method comprises receiving a message comprising an identifier for an address template, using the identifier to select the address template from a set of address templates, determining a set of memory addresses for a corresponding set of memory operations using the address template, and executing the memory operations.