Accelerator VM Live Migration Using Parallel DMA State Transfer

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing virtual machine migration technologies face challenges in efficiently transferring large amounts of state information, particularly for accelerators like GPUs, leading to noticeable performance drops during live migrations.

Innovation Solution

Utilizing direct memory access (DMA) operations by accelerators to store and restore state information in shared memory pages, parallelizing CPU and GPU operations to reduce latency and minimize the stop-and-copy phase duration.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If traditional CPU-based memory transfer is used to migrate accelerator state information, then the migration process is simple to implement, but the transfer speed is slow and latency is high

Engineering Contradiction:

Improvemigration speedVSAvoidmigration system complexity

Core Design Contradiction:

SpeedVSDevice complexity

Solution Approach 1:

The accelerator performs self-transfer of its state information to the destination host without requiring CPU intervention. The accelerator autonomously reads its state from local memory and writes it to the destination host's memory, eliminating the CPU bottleneck while maintaining implementation simplicity through accelerator-driven operation

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The patent replaces the traditional CPU-mediated mechanical memory transfer process with a direct accelerator-to-host memory transfer mechanism. This substitution eliminates the CPU involvement in the actual data transfer, significantly improving speed while the complexity is managed through hardware-level optimization rather than software complexity

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

2Reliability

If the stop-and-copy phase is used to ensure data consistency during migration, then data integrity is maintained, but the migration duration increases and performance drops

Engineering Contradiction:

Improvedata consistencyVSAvoidmigration duration

Core Design Contradiction:

ReliabilityVSLoss of time

Solution Approach 1:

The accelerator completes its state transfer to the destination host before the actual VM migration stop-and-copy phase begins. This preliminary action ensures that when the VM is stopped for copying, all accelerator state data is already available at the destination, eliminating the need to wait for state transfer during the critical consistency window and reducing overall migration duration

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The accelerator state transfer operation continues in the background during the VM's operational period, overlapping with other VM tasks. This continuous background transfer ensures data consistency is prepared beforehand while the VM remains running, avoiding the performance drop that would result from stopping the VM for state transfer

Inventive Principle:

Principle #20Continuity of useful action

3Quantity of substance

If large amounts of state information are transferred during migration, then complete VM state is migrated, but the transfer time increases and bandwidth requirements increase

Engineering Contradiction:

Improvestate information volumeVSAvoidtransfer time

Core Design Contradiction:

Quantity of substanceVSLoss of time

Solution Approach 1:

The accelerator autonomously manages its own state information transfer, directly reading from its local memory and writing to the destination host's memory without CPU mediation. This self-service approach maximizes transfer bandwidth and minimizes transfer time for large state datasets by eliminating CPU overhead and utilizing the accelerator's optimized memory interfaces

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The patent substitutes the CPU-based sequential memory transfer mechanism with a direct accelerator-driven parallel transfer system. This replacement enables high-speed transfer of large state information volumes by utilizing the accelerator's dedicated memory bandwidth and avoiding the CPU's memory bus bottleneck, significantly reducing transfer time for large datasets

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Applied Scientific Principles

This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.

Function Achieved in This Case

Significantly reduces the latency and duration of virtual machine migrations by maximizing bandwidth usage and minimizing CPU load, ensuring minimal performance impact on applications.

Implementation Method 1

the accelerator then performs a direct memory access (DMA) operation to save the first state information in the first migration buffer

Methodology Applied

Scientific EffectDirect Memory Access (DMA):

Data Source

PatentUS20260023587A1System and method for parallelized live migration of virtual machines that use accelerators

Publication Date: 2026.01.22 VMWARE INC

AI summary

A source host computer including a central processing unit (CPU), an accelerator, and memory, is configured to migrate a virtual machine (VM) that uses the accelerator to a destination host computer, by performing the steps of: requesting a driver of the accelerator to save first state information associated with the VM in a first migration buffer of the memory that is accessible to both the CPU and the accelerator, wherein the accelerator then performs a direct memory access (DMA) operation to save the first state information in the first migration buffer; extracting, using the CPU, the first state information from the first migration buffer; and transmitting the extracted first state information to the destination host computer.