Clock Interconnection Network for Sub-Microsecond Synchronization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
High-performance computing clusters face challenges in synchronizing internal clocks of nodes with precision due to variable transmission times in existing networks, leading to coherence issues and potential delays in timestamp information dissemination.
Innovation Solution
The method involves generating timestamp information by a source node and transmitting it through a clock interconnection network with adjusted transmission segments to ensure a constant time delay, allowing each node to accurately set its internal clock based on the information and the number of segments crossed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If timestamp information is transmitted through the Ethernet administration network using NTP, then all nodes can be synchronized, but the transmission precision is limited to about 100 microseconds due to variable transmission times
Solution Approach 1:
The patent divides the transmission path into multiple clock interconnection network segments between nodes, with each segment having a deterministic transmission time. This segmentation allows precise calculation of total transmission delay by summing individual segment delays, achieving sub-microsecond synchronization precision compared to the 100 microsecond limitation of Ethernet NTP.
Solution Approach 2:
The patent introduces a dedicated clock interconnection network as an intermediary between nodes for timestamp transmission. This separate communication channel is specifically designed for time synchronization, isolating it from the variable traffic of the Ethernet administration network and calculation data management network, thereby eliminating their transmission time variations.
2Speed
If the calculation data management network is used for timestamp transmission, then transmission speed is very fast (less than 1 microsecond), but the network experiences congestion during critical calculation phases
Solution Approach 1:
The patent creates a separate clock interconnection network segmented from the calculation data management network, dedicating specific transmission segments exclusively for timestamp synchronization. This prevents calculation traffic congestion from affecting time synchronization reliability while maintaining fast transmission speeds through optimized segment design.
Solution Approach 2:
The patent introduces a dedicated clock interconnection network as an intermediary channel that handles timestamp transmission separately from the calculation data management network. This intermediary network ensures reliable timestamp delivery during critical calculation phases by isolating synchronization traffic from computation traffic.
3Loss of time
If a star topology with all nodes directly connected to the source is used for clock interconnection, then transmission time is minimized and simplified, but the physical layout and propagation time dispersions in I/O components make this impractical
Solution Approach 1:
The patent applies local quality by optimizing each transmission segment individually with deterministic timing characteristics. Rather than requiring all nodes to be equidistant from the source (ideal star topology), each segment's transmission time is measured and compensated locally, allowing practical mesh-like topologies to achieve the same synchronization precision.
Solution Approach 2:
The patent changes the approach from geometric optimization (physical star topology) to parameter-based optimization (measuring and compensating transmission time for each segment). By making transmission time a configurable parameter rather than a fixed physical constraint, the system achieves minimal effective transmission time regardless of physical layout.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
The invention relates to a method for synchronising a server cluster comprising a plurality of nodes (Ni,j,k,l) that are each provided with an internal clock and are interconnected by a clock interconnection network comprising a plurality of transmission segments. Said method comprises the following steps: generation (100) of a timestamp information (H) by the internal clock of a source (N1,1,1,1) selected from the nodes (Ni,j,k,l) of the server cluster; transmission (110) of said timestamp information (H) from the source (?1,1,1,1) to the set of nodes (Ni,j,k,l) of the server cluster; and adjustment (130) of the internal clock of each node (Ni,j,k,l) of the server cluster on the basis of said timestamp information (H). Said method also comprises the following steps: adjustment (110) of the time of transmission of the timestamp information (H) by each transmission segment to a constant value established for each transmission segment; and once the timestamp (H) information has been received (120) by any one of the nodes (Ni,j,k,l) of the server cluster, adjustment (130) of the internal clock thereof on the basis of the timestamp information (H) and information relative to the transmission segments crossed between the source (?1,1,1,1) and said node (Ni,j,k,l).