RDMA Object Pool Atomic Access via Autonomous Manager

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current RDMA technologies support only limited atomic operations like Fetch-and-add and Compare-and-swap, which are insufficient for implementing complex algorithms such as selecting the next integer from a pool, and require multiple round trips for operations like accessing a shared message queue, leading to inefficiencies in high-performance computing systems.

Innovation Solution

The extension of RDMA transport to enable atomic access to a generic object pool, allowing fully offloaded operations through network interface controllers without host processor involvement, supporting operations like atomic get, push, and queue management in FIFO or LIFO structures, ensuring atomicity even with multiple agents accessing the same data structure.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If traditional RDMA atomic operations (Fetch-and-add, Compare-and-swap) are used, then data integrity is maintained, but operation complexity increases and performance decreases for complex algorithms

Engineering Contradiction:
Improveoperation efficiencyVSAvoidalgorithm complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The object pool manager autonomously manages the object pool without requiring host processor involvement. It handles allocation, deallocation, and atomic operations independently, allowing RDMA operations to be performed directly on the object pool with full atomicity guarantees, thereby simplifying algorithms and improving performance

Inventive Principle:
Principle #25Self-service

2Productivity

If multiple round trips are used for shared message queue access, then data integrity is ensured, but communication overhead increases

Engineering Contradiction:
Improvecommunication efficiencyVSAvoidcommunication time
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The patent merges the queue management functionality directly into the object pool, allowing shared message queue operations to be performed as single atomic operations on the object pool. This eliminates the need for multiple separate round trips while maintaining data integrity through the object pool manager's atomic access control

Inventive Principle:
Principle #5Merging (Combining)

3Productivity

If host processor involvement is required for object pool access, then operation flexibility is maintained, but system performance decreases

Engineering Contradiction:
Improveprocessing throughputVSAvoidoffload capability
Core Design Contradiction:
ProductivityVSExtent of automation

Solution Approach 1:

The patent extracts the object pool management functionality from the host processor and implements it as a separate, autonomous object pool manager. This allows object pool operations to be performed independently by the manager with full atomicity support, enabling complete offloading of these operations from the host processor and significantly improving processing throughput

Inventive Principle:
Principle #2Taking out (Extraction)

Data Source

PatentEP3217292B1Atomic access to object pool over RDMA transport network
Publication Date: 2020.10.14 MELLANOX TECHNOLOGIES LTD(IL)
  • EP3217292B1 patent drawingFigure 1
  • EP3217292B1 patent drawingFigure 2
  • EP3217292B1 patent drawingFigure 3~4

AI summary

In a fabric of network elements one network element has an object pool (60) to be accessed stored in its memory. A request for atomic access to the object pool (60) by another network element is carried out by transmitting the request through the fabric to the one network element, performing a remote direct memory access to a designated member of the object pool (60), atomically executing the request, and returning a result of the execution of the request through the fabric to the other network element.