Firefly：面向数据中心的可扩展、超精确时钟同步方案.pdf

1、Yuliang LiGoogle LLCFirefly:Scalable,Ultra-Accurate Clock Synchronization for DatacentersFirefly:Scalable,Ultra-Accurate Clock Synchronization for DatacentersYuliang LiGoogle LLCTIME APPLIANCES(TAP)Demand for Clock Accuracy is Growing TightermssnsLoggingDC Network TelemetryCongestion controlDistribu

2、ted databaseML schedulingFinancial ExchangeRDMASub-10ns Clock Sync For a Fair ExchangeExchange operatorParticipantsAccurate clock sync at sub-10ns is critical for ensuring fairness and market integritysimultaneous arrival at different machinesBuilding Block of Clock SyncIf dAB=dBAoffset=(RB-TA+TB-RA

3、)/2TARBdABClock AClock BTBRAoffset=RB-TA-dABdBAoffset=TB-RA+dBABuilding Block of Clock SyncTimeClock B-Clock AoffsetRB-TATB-RAIf dAB=dBAoffset=(RB-TA+TB-RA)/2Challenge to Sub-10ns Sync:AsymmetryTimeClock B-Clock AoffsetRB-TATB-RAAsymmetryChallenge to sub-10ns sync:Path-level asymmetry:different cabl

4、e lengths in forward and backward paths100s of ns,or10s of ns even w/equal nominal length cablesComponent asymmetry:delay of the same component is asymmetricOptical transceiver:up to 5ns/moduleSwitch,Challenge to Sub-10ns Sync:JitterTimeClock B-Clock AJitterRB-TATB-RAoffsetChallenge to sub-10ns sync

5、:Queuing delay10s to 100s of s,or10s to 100s of ns even w/strict prioritizationTimestamping jitter:up to 10ns per timestampChallenge to Sub-10ns Sync:DriftTimeClock B-Clock ARB-TATB-RAoffsetDynamic driftStatic drift:10s of ppm(s/s)Challenge to sub-10ns sync:Dynamic drift:10ppb(ns/s)in one secChallen

6、ge to Sub-10ns Sync:Errors in Time ServerTime server(s)CPUNICs jitters jitterRest of data centerSome sources of time themselves could have 100s of ns to s error!Near static error:10s to 100s of nsPCIePCIeCentralized architecture(PTP one-all,Huygens):Accuracy limited due to scalability of the central