Virtio Onload Interrupt Moderation for Low-Latency VM Data Paths
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing interrupt moderation schemes in computing environments, particularly in virtual machines, suffer from inefficiencies such as asynchronous updates leading to latency, race conditions causing false interrupts, and the need for polling across multiple VMs, which affects packet processing and increases latency.
Innovation Solution
Implement a Virtio standard system with a virtualizer interrupt moderation library that monitors current interrupt values and packet statistics, allowing dynamic adjustment of interrupt rates through an accelerator driver, thereby reducing the need for kernel involvement and eliminating race conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If interrupt moderation parameters are updated asynchronously in host memory, then device memory updates are delayed, but this causes the device to be unaware of parameter changes and increases latency
Solution Approach 1:
The patent introduces a shared memory region as an intermediary between host and device, with specific registers (INT_MOD_EN, INT_MOD_NUM, INT_MOD_IDX) that allow synchronized parameter updates. The device can atomically read these registers to obtain consistent interrupt moderation parameters without relying on asynchronous memory updates, thus reducing latency while maintaining parameter consistency.
Solution Approach 2:
The host writes interrupt moderation parameters to predetermined registers in the device before the device needs them. This preliminary action ensures that when the device reads the parameters, they are already available and consistent, eliminating the latency caused by waiting for asynchronous memory updates.
2Reliability
If device periodically reads parameters over PCI fabric, then device can obtain parameter values, but this slows interrupt generation process and affects packet processing flow
Solution Approach 1:
The patent replaces the mechanical PCI fabric communication with direct memory-mapped I/O to shared memory. Instead of periodic reads over the slow PCI fabric, the device can efficiently access interrupt moderation parameters through direct memory access to the shared memory region, significantly improving interrupt generation speed while maintaining parameter availability.
3Adaptability or versatility
If interrupt moderation is implemented in guest drivers, then interrupt rates can be controlled, but race conditions may cause false interrupts or interrupt delays
Solution Approach 1:
The patent implements interrupt moderation directly in the device firmware, making the device self-sufficient in managing its own interrupt rates. The device autonomously reads parameters from shared memory registers and applies them without requiring guest driver intervention, eliminating race conditions while maintaining the flexibility to adapt interrupt rates based on workload.
Solution Approach 2:
The shared memory registers act as an intermediary that provides a race-condition-free mechanism for parameter communication. The host writes to these registers, and the device reads them atomically, ensuring consistent parameter values without the race conditions that occur when multiple software components try to modify parameters simultaneously.
4Reliability
If driver notifications are disabled, then interrupt moderation can be enforced, but device must poll thousands of indices across hundreds to thousands of VMs, increasing latency
Solution Approach 1:
Instead of requiring the device to poll all possible indices across all VMs, the patent uses interrupt moderation parameters to limit the device's interrupt handling to only the necessary subset of work items. The INT_MOD_NUM register specifies the number of work items to process, and INT_MOD_IDX specifies the starting index, allowing the device to process only the required portion without excessive polling.
Solution Approach 2:
The host preliminarily determines and configures the interrupt moderation parameters (number of work items, starting index) before the device needs to process interrupts. This preliminary configuration allows the device to immediately begin processing the correct subset of work items without needing to poll through unnecessary indices, reducing polling latency while maintaining reliable interrupt moderation.
Data Source
AI summary
Systems and methods herein are for use with a Virtio standard and include at least one circuit to host a virtualizer component that enables different virtual machines (VMs), an accelerator driver, and a virtualizer interrupt moderation library, where the accelerator driver can be associated with an accelerator device. The accelerator driver can monitor interrupts associated with the different VMs to determine at least a current interrupt value and packet statistics that may be used by the virtualizer interrupt moderation library as a basis to provide an intended interrupt value to be used by the accelerator device for managing further interrupts to the different VMs.