精确测量电缆长度优化数据中心网络.pdf

1、Jasmeet Bagga,Software Engineer,MetaMehak Mahajan,Senior Director of Engineering,BroadcomCable Length MeasurementCable Length MeasurementJasmeet Bagga,Software Engineer,MetaMehak Mahajan,Senior Director of Engineering,BroadcomNETWORKINGData center staging velocity and how things go wrongImportance o

2、f cable length measurement in DSF,NSFDeployment user storyFeature implementation workflowSAI EnhancementsCall to ActionAgendaLarger AI clusters longer cables Buffer provisioning needs to be accurateRelevance of buffer provisioning in AI deploymentsUnder provisioning the buffer will lead to packet dr

3、opsOver provisioning the buffer will lead to higher latency Renewed focus on cable length measurement in AI NetworksData center builds in the age of AIPrometheus:1GW+cluster by 2026Prometheus:1GW+cluster by 2026Hyperion:5GW over next few yearsHyperion:5GW over next few yearsData centers-build them b

4、igger,faster Data center fiber design is a complex processInitial fiber modeling and projectionOrderingExecution-fiber layoutWith scale and speed chances of error go upData center builds in age of AI-Bigger,faster Enter DSFDSF is at the heart of Prometheus buildsDSF has strict requirements forEnd-to

5、-end cable lengthVOQ Switch(edge/leaf)Fabric switch cable length Requirements become stricter with 2-stage DSFHyperport-3.2T/1.6T DCI portStrict requirements on inter-member port skewEnter DSFFor large clusters fabrics connect to each other over multi-km linksImportant to get distance correct for op

6、timal PFC,congestion control settingsNot just DSFActual deployments didnt go as plannedAccelerated schedules meant not all requirements were met Requirements got stricter due to error-detection settingsCable length monitoring detected we we