设计更优秀的 NVMe 控制器的经验.pdf

1、Matt Bromage ARMTim Amundsen-ScaleFluxExperiences in Designing a Better NVMe ControllerExperiences in Designing a Better NVMe ControllerTim Amundsen-ScaleFluxMatt Bromage ArmSTORAGEThe OCP Storage specification has seen dramatic growth since 2009.The current version,is now 220 pages long(up from 56)

2、and includes over 475 requirements(up from 175).Growing Requirements on Data Center SSDs150200250300350400450500201820192020202120222023202420252026OCP Storage Specification Requirements over TimeOCP Storage Specification Requirements over TimeProblem Statement:NVMe SOCs Struggle to Keep UpMounting

3、Demands:-More Features:Security,Manageability,etc.-More Performance:PCIe Gen,IOPs for AI,etc.-More Capacity:128+TB in 2026NVMe SoCs face significant design challenges to meet increasing demands within tight and resolute space,power,and cost constraints.Arm A53 Cores powered the SFX3000 and FX5016 So

4、CsGen 4 and 5 performanceOCP 2.5 and 2.6 complianceTransparent CompressionEtc.ScaleFlux SoC Designs and Transition to R82Arm R82 Cores drive success for Gen 6 and its forerunner,the FC6116 and FC5116 SoCsSupports larger memory via 64-bit architectureAchieves lower latency and higher throughput with

5、real-time processingOptimizes efficiency with single-cycle access to Tightly Coupled MemoryHuge per-core real-time performance jump:81%over R5 and 36%over R8(CoreMark/MHz)Increased physical addressing bits from 32 to 40 to support onboard DRAM sizes up to 1TB(previous limit was 4GB!)Increased SMP su

6、pport for up to 8 cores per cluster as well as a new addition of a shared L2 cache for multicore shared memoryCHI coherencyfor advanced CXL controllers Near-Data Computecapabilities including MMU(opt.)to run native Linux+Neon ML engine(opt.)Arm Cortex-R82Highest performance,fully-featured Cortex-R f