Gateway Sync Group Management for Data Consistency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In complex data processing systems, particularly in machine intelligence applications, there is a challenge in synchronizing data exchange between accelerator tiles and gateways, as well as between multiple accelerators, which affects the efficiency of parallel processing and data consistency.
Innovation Solution
A gateway system is introduced that includes synchronization propagation circuits to manage and control synchronization requests and acknowledgments, allowing for flexible definition of sync groups and credit-based data exchange mechanisms to ensure data consistency across accelerators and external storage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If barrier synchronization is implemented between tiles and gateways, then data consistency is ensured, but processing latency increases
Solution Approach 1:
The system divides the synchronization mechanism into two distinct segments: barrier synchronization for intra-accelerator tile communication and credit-based synchronization for accelerator-to-gateway communication. This segmentation allows each synchronization type to be optimized independently, reducing the impact of barrier synchronization on overall processing latency while maintaining data consistency where required.
Solution Approach 2:
A credit-based intermediary mechanism is introduced between accelerators and gateways to manage data exchange without requiring full barrier synchronization. The credit system acts as a mediator that tracks data availability and enables asynchronous progress, reducing latency while ensuring consistency through credit validation before data transfer.
2Adaptability or versatility
If multiple accelerators are connected to different gateways, then system scalability is improved, but synchronization control complexity increases
Solution Approach 1:
The credit-based synchronization mechanism is designed as a universal interface that works across multiple accelerator-gateway connections. The same credit tracking and validation logic applies regardless of which gateway or accelerator is involved, providing a consistent synchronization model that scales to multiple accelerators connected to different gateways without increasing control complexity.
Solution Approach 2:
The system uses credit counts as adjustable parameters to control the degree of synchronization required. By changing credit parameters (such as credit thresholds and allocation rates), the system can adapt synchronization behavior to match different scalability scenarios, from single-accelerator to multi-accelerator configurations, without changing the fundamental synchronization control mechanism.
3Productivity
If credit-based data exchange is implemented, then processing throughput is improved, but synchronization management complexity increases
Solution Approach 1:
The credit-based system implements self-service synchronization management where the gateway automatically tracks credit allocation, validates data exchange requests, and updates synchronization state without external intervention. This automation reduces the apparent complexity for users while maintaining high throughput through efficient credit-based flow control and data exchange coordination.
Data Source
AI summary
A work accelerator is connected to a gateway. The gateway enables the transfer of data to the work accelerator from an external storage at pre-compiled data synchronisation points attained by the work accelerator. The work accelerator is configured to send to a register of the gateway an indication of a sync group comprising the gateway. The work accelerator then sends to the gateway, a synchronisation request for a synchronisation to be performed at an upcoming pre-compiled data exchange synchronisation point. The sync propagation circuits are each configured to receive at least one synchronisation request and propagate or acknowledge the synchronisation request in dependence upon the indication of the sync group received from the work accelerator.