PDF 原檔：報告_OCP2026_Meta_CXL超大規模部署_20261015_original.pdf
講者公開簡報（OCP 官方 Recorded Content 頁提供下載）。文字以 PyMuPDF 逐頁抽取，
### p<n>為頁碼；圖表內容見原檔。
圖片清單（已驗證 2026-10-19）
| 檔名 | 頁 | 分類 | 親眼所見內容 |
|---|---|---|---|
報告_OCP2026_Meta_CXL超大規模部署_20261015_p10.png |
p10 | 真資料圖 | CXL 記憶體擴充 ASIC 方塊圖：Host CPU 經 CXL/PCIe 連 CXL Subsystem（PCIe PHY＋CXL Controller），經 CXL.mem 接 CXL2DDR Bridge 再接 DDR Subsystem（兩組 MEMC＋PHY）連到 DIMM；下方 Control Plane 含 Boot／Control／Security CPU、SRAM、低速周邊（SMBus/QSPI/UART/JTAG） |
報告_OCP2026_Meta_CXL超大規模部署_20261015_p15.png |
p15 | 真資料圖 | Meta 導入 CXL 的工作負載效益表：Caching 吞吐 +25%、保留 5–10x；Distributed Caching 平均查詢延遲 -29%；Data Warehouse 每台 executor +25%；DevInfra 每台 VM +33%、機隊 -15%；ML Parameter Server 伺服器 -25%、吞吐 +12% |
原始內容
p1
Markandeya Mudigonda Production Systems Engineer, Meta Vineet Parekh Production Systems Engineer, Meta CXL in Hyperscale Environment Fleet-Scale Challenges, Failure Modes & the Path to Better Debuggability SERVER: COMPOSABLE MEMORY SYSTEMS (CMS)
p2
Failure Detection in Fleet
p3
CXL At-Scale Telemetry Collection SEL event logs: BMC system event logs, hardware events with timestamps DMESG logs: kernel messages: AER errors, MCEs, CXL driver events Crash dumps: captured when a fatal error takes down a process or the system Device health counters: device telemetry: correctable error counts, temperature, media health
p4
Fleet Failure Debug Overview
p5
Hardware Remediation @ Scale Failure Detection – MachineChecker ● Runs hardware checks periodically ● Host ping, memory, CPU, NIC, dmesg, S.M.A.R.T., power supply, SEL, etc.
p6
Hardware Remediation @ Scale Failure Detection – MachineChecker Failure Digestion – Unified Repair ● Auto Remediation ● Picks up hardware failures, processes logged information, and executes custom-made remediation accordingly
p7
Hardware Remediation @ Scale Failure Detection – MachineChecker Failure Digestion – Unified Repair Low-Level Software Fix – Ticketing ● Handles low-level software fixes such as firmware update and reimaging
p8
Hardware Remediation @ Scale Failure Detection – MachineChecker Failure Digestion – Unified Repair Low-Level Software Fix – Ticketing Manual Fix – Repair Ticketing ● Creates repair tickets for DC technicians to swap CXL ASIC / DIMMs ● Provides detailed logs throughout the auto-remediation ● Logs repair actions for further analysis
p9
CXL Failures in Fleet
p10
CXL Memory Expander Failure Domains
p11
ASIC Enumeration Issues ROOT CAUSES DETECTION CXL FW self-test failure SPI Flash Failure Link Training Failure ASIC Init failure CXL Link Never Trains PCIe scan (In-band) BMC polls ASIC heartbeat (OOB)
p12
Run time failures ROOT CAUSES DETECTION PCIe Errors Surprise link down, completion timeout CXL Viral Error Memory CE Storm ASIC Unresponsive Host Gets MCE Fatal error → MCE (In-band) BMC detects heartbeat loss (OOB)
p13
CXL Memory Failures TYPES OF MEMORY ERRORS Correctable ECC Error (two Symbol Error) Uncorrectable ECC Error (Multi-bit Error) DDR Training Failures (Margin Issues) ROOT CAUSES DETECTION ● Defective DIMM ● SI/Power/Thermal issues System Event Log (OOB)
p14
One Symptom. Several Possible Causes. INSIDE THE DEVICE — NOT OBSERVABLE DIMM fault Controller/ logic bug Firmware issue observability boundary WHAT THE HOST SEES CPU → CXL.mem Read Timeout CONSEQUENCE Unmaskable MCE — host lost without warning All in-flight work destroyed
p15
Why CXL: Workload Impact at Scale Category Primary Bottleneck Key Improvement Caching Memory Capacity Throughput +25%, Retention 5–10x Distributed Caching Network BW Avg query latency −29% Data Warehouse Alloc Mem / CPU Executors/server +25% DevInfra (Devmachines) Memory Capacity VMs/server +33%, Fleet −15% ML Parameter Server Memory Capacity Servers −25%, Throughput +12%
p16
Call to Action Problem to solve ●A CXL memory expander device can be fully spec-compliant and still be non debuggable at fleet scale What should be addressed? ●Define and standardize requirements for CXL device debuggability and telemetry at Fleet Scale ●Keen to collaborate with the ecosystem and CXL Consortium
p17
Thank You! Click here to review OCP Trademark Guidelines
p18
MARKET IMPACT & ADOPTION ] Please use the appropriate Global Track name and icon. AI OPEN DATA CENTER AI COMPUTING CONTINUUM AI CLUSTERS COOLING ENVIRONMENTS AI OPEN DATA CENTER: Power Distribution COOLING: COLD PLATE COOLING: IMMERSION ] OPEN PLATFORM FIRMWARE (OPF) NETWORKING
p19
SYSTEMS MANAGEMENT TEST & VALIDATION OCP SPECIAL FOCUS: PHOTONICS TIME APPLIANCES (TAP) Please use the appropriate Global Track name and icon. SECURITY SERVER: COMPOSABLE MEMORY SYSTEMS (CMS) SERVER: OPEN CHIPLET ECONOMY SUSTAINABILITY STORAGE ] RACK & POWER
p20
Please use one of these membership logos to designate your company’s membership level.
p21
Please use this logo if you are hosting an area in the Innovation Village. Please use one of these logos if your facility is an OCP Ready facility or you’re hosting an Experience Center. Please use one of these logos if your product has been recognized as OCP Accepted TM Please use this logo if your product has been approved through the OCP S.A.F.E. TM program. Please use this logo if you are an OCP Solution Provider.
p22
Please use the appropriate icon representing the Regional Project Group.