Latin Square Fat-Tree Switch Allocation for Parallel Communication
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In parallel computer systems using a Latin square fat-tree topology, all-to-all communication often results in route conflicts due to multiple packets being transmitted simultaneously over the same path, leading to increased communication time and inefficiencies.
Innovation Solution
The system employs a management apparatus that allocates leaf switches to jobs based on a finite projective plane lattice structure, generating communication tables to ensure that each job uses distinct paths, avoiding route conflicts by utilizing regular and static routing in the InfiniBand network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If multiple jobs perform all-to-all communication concurrently in a Latin square fat-tree system, then system throughput and productivity are improved, but route conflicts occur causing packet transmission delays and communication inefficiency
Solution Approach 1:
The patent segments the leaf switches into multiple groups, where each group is assigned to a specific job. This segmentation ensures that packets from different jobs traverse different paths through the spine switches, eliminating route conflicts while maintaining concurrent execution. The lattice part of the finite projective plane is divided into multiple columns, with each column corresponding to a leaf switch group allocated to a particular job.
2Device complexity
If a Latin square fat-tree topology is used to couple many servers with fewer switches, then construction cost is reduced, but route conflicts occur during all-to-all communication
Solution Approach 1:
The patent applies preliminary action by pre-establishing a routing schedule based on the finite projective plane geometry before communication occurs. The management apparatus generates communication tables that dictate the sequence and paths of packet transmissions, ensuring that route conflicts are prevented in advance rather than resolved during communication. This scheduling approach allows the system to maintain the cost-effective Latin square fat-tree topology while eliminating route conflicts.
Data Source
AI summary
A system includes spine switches, leaf switches, information processing apparatuses, and a processor configured to allocate a first leaf switch group to a first job, the first leaf switch group corresponding to a first column in a lattice part including points other than points at infinity of a finite projective plane corresponding to a Latin square fat-tree, and allocate a second leaf switch group to a second job, the second leaf switch group corresponding a second column, and transmit first schedule information on first communication related to the first job to a first information processing apparatus coupled to the first leaf switch group, and transmit second schedule information on second communication related to the second job to a second information processing apparatus coupled to the second leaf switch group, wherein the first and second communication are collective communication in which each of the information processing apparatuses communicates with others.