Turning PCIe Hints into Cache Hits: Enabling Smart Data Cache Injection in Linux
summary
AMD’s Smart Data Cache Injection (SDCI uses PCIe TLP Processing Hints to steer inbound DMA into the consumer CPU’s L2 cache; Linux adds a new ACPI _DSM interface and TPH API so drivers can query steering tags and program steering tag entries. Tests on an AMD Tyrion Zen 5 processor with a Broadcom P2100 100G NIC show about ~70% memory bandwidth reduction in 1‑queue setups and >50% in 4‑queue, plus consistent latency improvement across payloads; patches land in kernel 6.13/6.15/6.18.
AI-written summary · about
chapters (6)
- 0:00Introduction to SDCI
- 1:03What is AMD SDCI
- 3:59Kernel integration and API
- 6:10Performance improvements
- 9:06Latency benefits
- 10:08Current status and future
related talks
01h 06mPerformance engineering: being friendly to your hardware
19mP2CACHE: Exploring Tiered Memory for In-Kernel File Systems Caching
42mExtending Container Performance Isolation: Regulating Memory Bandwidth & Cache in the Kernel
29mImprove Xeon IRQ throughput with posted interrupt
26mAnalysis of AMD HW-assisted vIOMMU Implementation and Performance
01h 05mPerformance Optimization in Software Development - Being Friendly to Your Hardware
33mCPU Hardware Counters
22mAccelerate the Image Distribution for DADI Using Intel DSA