Improve Xeon IRQ throughput with posted interrupt
summary
Enables posted MSI on bare metal and adds a kernel RFC that coalesces delivery so only the leading interrupt in a burst notifies the CPU, avoiding slow APIC notification paths. FIO on Samsung Gen5 NVMe shows per-socket throughput falling from ~2 million IOPS per drive at two drives to 1.1M at four and <50% at eight; the patch recovers to line rate (~2M). Tests also use the U built-in data streaming accelerator to generate millions of IRQs; implementation is Intel x86 kernel/IOMMU changes with a three-loop default to bound polling.
AI-written summary · about
chapters (8)
- 0:00Introduction and motivation
- 1:04Interrupt performance problem
- 6:30Proposed solution overview
- 9:05Posted interrupt mechanics
- 15:12Performance improvements
- 17:20Implementation details
- 24:18Q&A and audience feedback
- 28:36Summary and closing
related talks
31mMeasuring the Impacts of the Preempt-RT Patch
09mInterrupts: The Hidden World of Linux Performance
26mAnalysis of AMD HW-assisted vIOMMU Implementation and Performance
21mNVMe Emulation Performance Optimization
16mTurning PCIe Hints into Cache Hits: Enabling Smart Data Cache Injection in Linux
19mAsynchronous I/O Stack: A Low-latency Kernel I/O Stack for Ultra-Low Latency SSDs
30mOptimizing vLLM for Intel CPUs and XPUs
19mSpeeding Up VM’s I/O Sharing Host's io_uring Queues With Guests