FPGA Cloud Platform Far-End Data Migration via Custom iRDMA
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current FPGA cloud platforms face challenges in efficiently migrating data between FPGA boards in a Just a Bunch Of FPGAs (JBOF) topology, as existing RDMA protocols are complex and resource-intensive, lacking a standardized method for data transmission between FPGA boards.
Innovation Solution
A far-end data migration device and method utilizing an FPGA cloud platform, featuring a SHELL and FPGA Accelerator Unit (FAU) with an customized iRDMA, MAC, and PCIE interfaces, employing LZ77 and Huffman algorithms for data compression and migration between FPGA acceleration cards, simplifying data migration and reducing resource usage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If RDMA protocol standard is used for data transmission between FPGA boards, then data migration function can be realized, but device complexity and resource occupancy increase significantly
Solution Approach 1:
The patent extracts the essential data migration function from the complex RDMA protocol standard, implementing a simplified custom protocol that retains core functionality while removing unnecessary complexity. The custom protocol includes only the necessary components for data transmission between FPGA boards, eliminating the need to implement the full RDMA stack.
Solution Approach 2:
The patent uses lightweight, disposable protocol messages for data migration instead of maintaining complex stateful connections. The custom protocol employs simple request-response message structures that can be quickly sent and processed, reducing the computational burden and resource occupancy compared to persistent RDMA connections.
2Adaptability or versatility
If RDMA protocol standard is implemented, then data transmission between hosts is enabled, but FPGA resource occupancy increases
Solution Approach 1:
The patent extracts only the essential data transmission capability from RDMA, implementing a minimal custom protocol that uses a small portion of FPGA resources. The custom protocol implements data migration through simple message passing mechanisms that require far fewer logic resources than the full RDMA stack.
3Productivity
If customized iRDMA is used for data migration between FPGA acceleration cards, then transmission efficiency is improved, but protocol complexity increases
Solution Approach 1:
The patent applies local quality optimization by implementing protocol-specific features only where needed for high-performance data migration between FPGA cards. The custom iRDMA protocol includes localized optimizations such as direct memory access capabilities and customized error handling that improve transmission efficiency without requiring complex protocol features throughout the entire system.
Data Source
AI summary
A far-end data migration device and method based on a FPGA cloud platform. The device includes a server, a switch, and a plurality of FPGA acceleration cards. The server transmits data to be accelerated to the FPGA acceleration cards by means of the switch. The FPGA acceleration cards are configured to perform a primary and/or secondary acceleration on the data, and are configured to migrate the accelerated data. The method includes: transmitting data to be accelerated to a FPGA acceleration card from a server by means of a switch; performing, by the FPGA acceleration card, a primary and/or secondary acceleration on the data to be accelerated; and migrating, by the FPGA acceleration card, the accelerated data.