How Edge High‑Performance VMs Achieve Near‑Bare‑Metal Performance

Why Edge Computing Needs New Performance Guarantees

AI, autonomous driving, cloud gaming, and live audio‑video services require faster data processing, lower network latency, reduced bandwidth costs, and stronger security. Traditional centralized clouds cannot fully meet these demands, prompting the rise of edge computing.

Volcano Engine Edge Nodes

Volcano Engine provides general‑purpose edge nodes across Chinese provinces and carriers. Built on a native edge‑cloud OS, they deliver elastic, reliable, distributed compute and low‑latency networking, enabling rapid deployment of services at the network edge.

Edge Compute Forms

Edge Virtual Machines : Elastic, stable, high‑performance VMs supporting x86, ARM, and GPU resources.

Edge Bare‑Metal : High‑performance, quickly provisioned bare‑metal servers.

The focus of this article is on performance optimization for edge VMs.

Design Goals of High‑Performance Edge VMs

Minimize VM Exit count to approach bare‑metal performance.

Guest images work without modification.

Support hot‑upgrade, live‑migration, and mixed deployment with regular VMs.

Enable coexistence with generic VM instances.

Core Architecture

The solution builds on QEMU + KVM for virtualization. While QEMU + KVM already leverages hardware acceleration, many sources of overhead remain, such as IPI, timer, and device interrupts. By modifying the host kernel and KVM module, a near‑zero‑loss hardware‑pass‑through design is achieved.

Key Techniques

Guest‑Interrupt No‑Exit : Configure VMCS INTR_EXITING so external interrupts are delivered directly to the guest without causing a VM Exit.

Timer Interrupt Passthrough : Assign LAPIC timers to the guest and disable VM Exits for TSC‑Deadline MSR operations.

VFIO Interrupt Passthrough : Rewrite IOMMU IRTE entries to deliver device interrupts straight to the guest, bypassing Posted‑Interrupt overhead.

IPI Fast‑Path : Rewrite VMM IPI handling in assembly to inject IPI interrupts directly into the guest, reducing exit latency.

Kernel Resource Isolation : Use dynamic NOHZ_full and isolCPU techniques to keep host timer and device interrupts away from vCPUs dedicated to high‑performance VMs.

Performance Evaluation

Micro‑benchmarks show dramatic reductions in IPI latency, timer latency, and MSR write overhead.

VM Exit Statistics indicate a >99% decrease in VM Exits for high‑performance instances compared with generic ones.

Cloud‑Server Benchmarks using wrk, ab, redis‑benchmark, and netperf demonstrate 6%‑16% overall performance gains.

Business Impact

CDN Scenario : CPU usage reduced by 13.9%‑23.2% versus generic VMs and within 0.2%‑2.9% of bare‑metal.

Audio/Video Live Streaming : 24.2% performance improvement over generic VMs, comparable to bare‑metal.

Acceleration Scenario : Near‑bare‑metal latency stability by eliminating VM‑Exit‑induced jitter.

Conclusion and Outlook

Edge high‑performance VMs are now live, delivering near‑bare‑metal performance with low virtualization overhead across multiple edge workloads. Future work will extend support to GPU and other heterogeneous accelerators, continuing to provide cost‑effective, stable compute for edge applications.