Mapping-Aware MPI Collectives for Topology-Optimized Data Transfer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional MPI libraries are agnostic to process-to-resource mappings and hardware topology, leading to suboptimal utilization of hardware resources and inefficient data transfer in collective operations.
Innovation Solution
The implementation of mapping-aware and memory topology-aware MPI collective communication logic, which considers process affinity and hardware topology to optimize data transfer and reduce latency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If conventional MPI libraries use agnostic communication mechanisms to all processes, then implementation simplicity is maintained, but data transfer efficiency deteriorates due to ignoring process-to-resource mappings and hardware topology
Solution Approach 1:
The patent segments the communicator into multiple affinity domains based on process-to-resource mappings and hardware topology. Each affinity domain is further divided into partitions, creating a hierarchical structure that enables targeted communication. This segmentation allows the system to move from generic all-to-all communication to specialized within-domain communication, improving data transfer efficiency while maintaining implementation clarity through structured organization.
Solution Approach 2:
The patent applies local quality by making communication behavior adaptive to local hardware characteristics. Each affinity domain is configured with specific communication parameters based on its underlying resource mappings and topology. This allows different parts of the system to use optimization strategies tailored to their local hardware environment, thereby improving overall data transfer efficiency without sacrificing implementation simplicity through the use of a unified framework.
2Reliability
If MPI collectives transfer data between all ranks irrespective of hardware topology, then communication completeness is ensured, but data movement overhead increases across affinity domain partitions
Solution Approach 1:
The patent segments the communication space into affinity domains and partitions, allowing data to be transferred only within relevant domains rather than across all possible rank pairs. This segmentation maintains communication completeness for required data exchanges while eliminating unnecessary cross-domain data movements, thereby reducing data movement overhead and energy consumption.
Solution Approach 2:
The patent extracts and identifies the essential communication requirements within each affinity domain, separating necessary data transfers from unnecessary ones. By taking out only the required communication operations and executing them within appropriate domains, the system ensures communication completeness for essential data while removing redundant data movement overhead across affinity domain partitions.
3Device complexity
If conventional MPI approaches use flat tree topology for collectives, then implementation simplicity is maintained, but communication latency increases due to ignoring process affinity and hardware topology
Solution Approach 1:
The patent introduces dynamic topology selection that adapts communication patterns based on process-to-resource mappings and hardware topology characteristics. Rather than using a static flat tree structure, the system dynamically determines optimal topologies for each affinity domain, allowing communication latency to be minimized according to actual hardware conditions while maintaining reasonable implementation complexity through parameterized configurations.
Solution Approach 2:
The patent changes key parameters of the communication topology based on hardware characteristics and process mappings. By adjusting topology parameters such as tree depth, branching factor, and domain hierarchy levels according to actual system configuration, the system optimizes communication latency without requiring fundamentally complex new structures, thus balancing performance improvement with implementation feasibility.
Data Source
AI summary
A message passing interface processing system is described. In accordance with message passing logic, a node selects an affinity domain for communication of data associated with a message passing interface and selects a first rank of a first process of the message passing interface assigned to a first partition of the affinity domain as a first partition leader rank and an affinity domain leader rank. The node selects a second rank of a second process of the message passing interface assigned to a second partition of the affinity domain as second partition leader rank, receives the data at the first partition leader rank, and communicates the data from the first partition leader rank to the second partition leader rank.