Coherent Attached Bus RDMA Key-Value Store

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional key-value (KV) stores tend to saturate the CPU before network bandwidth and require pinning and registering memory with RDMA adapters, leading to overhead and limited scalability, as well as storage limitations in add-on accelerators like FPGAs.

Innovation Solution

The use of a coherent attached bus (CAB) allows for direct and coherent access to shared memory by multiple RDMA adapters without the need for pinning and registering memory, enabling flexible memory management and increased scalability by allowing multiple adapters to access the same memory space.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If memory is pinned and registered with RDMA adapters for KV store access, then access reliability is improved, but device complexity and overhead increase

Engineering Contradiction:
Improveaccess reliabilityVSAvoidmemory management complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent introduces a CAB (Coherent Accelerator Processor Interface) bus as an intermediary between RDMA adapters and shared memory. This CAB bus provides the memory pinning and registration functionality centrally, allowing multiple RDMA adapters to access the same memory regions without each adapter needing its own private registration. The CAB bus acts as the mediator that manages the coherence and access rights, reducing the complexity at each adapter while maintaining reliable access.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If private memory registration is used for each RDMA adapter, then access security is improved, but scalability is limited

Engineering Contradiction:
Improveaccess securityVSAvoidscalability
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The patent implements a universal memory space that can be shared by multiple RDMA adapters through the CAB bus. Instead of each adapter having its own private registered memory regions, the system creates a universal shared memory space that all adapters can access. The CAB bus manages the security and coherence attributes, allowing the system to scale to many adapters while maintaining secure access control through a single unified interface.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Measurement precision

If CPU involvement is increased for memory management, then control precision is improved, but productivity decreases

Engineering Contradiction:
Improvecontrol precisionVSAvoidaccess bandwidth
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent enables RDMA adapters to perform self-service memory access through the CAB bus without requiring CPU intervention for each access operation. The adapters can directly access the shared memory regions managed by the CAB bus, performing read and write operations autonomously. This self-service capability eliminates the need for CPU involvement in memory management and data transfer, thereby maximizing access bandwidth and productivity while the CAB bus maintains control precision through its coherence management.

Inventive Principle:
Principle #25Self-service

4Speed

If add-on accelerators like FPGAs are used, then processing speed is improved, but storage capacity is limited

Engineering Contradiction:
Improveprocessing speedVSAvoidstorage capacity
Core Design Contradiction:
SpeedVSVolume of stationary object

Solution Approach 1:

The patent moves the storage capacity dimension from the accelerator card itself to the system's shared memory resources. Instead of being constrained by the limited onboard storage of FPGAs or other add-on accelerators, the system allows these high-speed processing units to access large容量的 shared memory through the CAB bus. This dimensional shift separates processing speed (provided by the accelerator) from storage capacity (provided by shared memory), allowing both to be optimized independently.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Data Source

PatentUS10225344B2High-performance key-value store using a coherent attached bus
Publication Date: 2019.03.05 INTERNATIONAL BUSINESS MACHINE CORPORATION
  • US10225344B2 patent drawing
  • US10225344B2 patent drawing
  • US10225344B2 patent drawing

AI summary

An infrastructure for servicing remote direct memory access (RDMA) requests without the need to pin memory and/or register memory for access by only a single RDMA adapter. Rather, RDMA requests can be coherently serviced from memory accessible to multiple adapters, and that memory can be flexibly managed (such as by swapping it out of main memory or moving it around within main memory). This coherent servicing of RDMA requests is achieved through the use of a coherent attached bus connecting the RDMA adapters to main memory that owns the address space.