Accelerator Scheduling Statistics Handling

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

In virtualized data processing systems, there is a latency issue in obtaining scheduling statistics from accelerators, which hinders efficient resource allocation and management, particularly when multiple virtual machines compete for shared accelerator resources.

Innovation Solution

The accelerator periodically writes scheduling statistics to a shared storage area without a request from the hypervisor, allowing the hypervisor to access this information quickly and efficiently, reducing latency and improving resource allocation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of time

If the hypervisor requests scheduling statistics from the accelerator when needed, then the hypervisor can obtain the required information, but the process introduces significant latency (approximately 100 μs) that hinders efficient resource allocation

Engineering Contradiction:

Improvelatency in obtaining scheduling statisticsVSAvoidefficiency of resource allocation

Core Design Contradiction:

Loss of timeVSProductivity

Solution Approach 1:

The accelerator proactively writes scheduling statistics to shared memory in advance, before the hypervisor needs to request them. This preliminary action eliminates the latency associated with on-demand requests, as the statistics are already available when the hypervisor needs to make scheduling decisions.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

Shared memory serves as an intermediary between the accelerator and the hypervisor. Instead of direct communication requiring request-response cycles, the accelerator writes to shared memory and the hypervisor reads from it, enabling asynchronous data exchange and eliminating request latency.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Loss of time

If the accelerator periodically writes scheduling statistics to shared memory without requests, then the hypervisor can access information quickly (reducing latency to approximately 100 ns), but the accelerator performs unnecessary writing operations

Engineering Contradiction:

Improvetime to access scheduling statisticsVSAvoidenergy consumed by unnecessary writing operations

Core Design Contradiction:

Loss of timeVSLoss of energy

Solution Approach 1:

The accelerator writes scheduling statistics to shared memory at periodic intervals rather than continuously or on every event. This periodic action balances the need for fresh data with the cost of writing operations, ensuring statistics are updated frequently enough to be useful while avoiding excessive energy consumption from continuous writing.

Inventive Principle:

Principle #19Periodic action

Data Source

PatentUS9678781B2Methods of and data processing systems for handling an accelerator's scheduling statistics

Publication Date: 2017.06.13 ARM LTD

AI summary

A data processing system comprises one or more processors that each execute one or more operating systems. Each operating system includes one or more applications. The system also comprises an accelerator that provides a shared resource for a plurality of the applications, an input/output module comprising one or more input/output interfaces for the submission of tasks to the accelerator, a hypervisor that manages the allocation of the input/output interfaces to the one or more operating systems and a storage area accessible by the hypervisor and the accelerator. The accelerator is capable of writing one or more selected pieces of information representative of one or more scheduling statistics of the accelerator periodically to the storage area without having received a request for the one or more selected pieces of information from the hypervisor.