Latin Square Fat Tree Network Topology for Parallel Communication
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In parallel computers, optimizing network topology to reduce construction costs and increase throughput is challenging, particularly in avoiding route conflicts during collective communications like Allreduce and all-to-all communications in fat tree structures.
Innovation Solution
A Latin square fat tree system is employed, where Spine and Leaf switches are connected in a specific configuration, using regular and fixed routing to avoid route conflicts by utilizing different Spine switches for different phases of collective communications, and a management apparatus generates communication tables to coordinate these communications.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If a Latin square fat tree topology is used to connect more servers with fewer switches, then construction cost is reduced and network density is improved, but route conflicts occur during parallel collective communications
Solution Approach 1:
The system segments the collective communication into multiple phases, with each phase using a different Spine switch. This segmentation allows parallel execution of communications that would otherwise conflict on the same network path, resolving the route conflict problem while maintaining the compact Latin square fat tree topology.
Solution Approach 2:
The system dynamically assigns different Spine switches to different phases of collective communication based on the communication pattern. This dynamic routing strategy adapts the network path selection to avoid conflicts, enabling efficient utilization of the limited switch resources in the Latin square fat tree.
2Productivity
If collective communications are executed in parallel to improve throughput, then productivity is increased, but route conflicts occur that reduce reliability
Solution Approach 1:
The system executes collective communications in periodic phases, where each phase uses a specific Spine switch. This periodic structure allows multiple communications to proceed in parallel without conflict by cycling through different Spine switches, thereby maintaining both high throughput and route conflict avoidance.
Solution Approach 2:
The system maintains continuous useful action by overlapping the execution of multiple collective communications across different phases. While one communication phase uses a particular Spine switch, another phase simultaneously uses a different Spine switch, ensuring continuous utilization of network resources without interruption or conflict.
3Device complexity
If a general two-stage fat tree is used for simplicity, then device complexity is reduced, but fewer servers can be connected with the same number of switches
Solution Approach 1:
The Latin square fat tree topology provides multi-functionality by enabling multiple collective communication patterns (Allreduce, all-to-all, etc.) to be executed efficiently on the same network structure. This universal design allows the network to serve more servers with the same number of switches compared to a general two-stage fat tree.
Solution Approach 2:
The system changes the topological parameters of the fat tree by using a Latin square structure instead of a general two-stage structure. This parameter change optimizes the network for higher server connectivity and efficient collective communications, achieving better performance without proportionally increasing the number of switches.
Data Source
AI summary
An information processing system includes Spine switches, Leaf switches coupled to the Spine switches in a form of a Latin square fat tree, and apparatuses each coupled to any one of the Leaf switches and including a processor. The processor performs, in a case where the processor is included in one of first apparatuses coupled to one of first Leaf switches, first collective communication with others of the first apparatuses on a route via a first Spine switch. The first Leaf switches correspond to at least a portion of points other than points at infinity of a finite projective plane corresponding to the Latin square fat tree. The processor performs second collective communication with others of the first apparatuses on a route via a second Spine switch at each phase of the first collective communication. The second Spine switch is different from the first Spine switch.