资源描述:
《【t112017-数据工程和技术分会场】物联网和人工智能领域内置芯片分析的意外之旅》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库。
1、Future-proofingBI:anunexpectedjourneytoleverage‘In-Chip’analyticsinIoTandAIAniManianHeadofProductStrategySIMPLIFYINGBusinessAnalyticsforCOMPLEXData“ThekeystrengthofSisenseistheplatform’scapabilitytoeasilyhandleandmanagelargeanddiversedatasets,andanalyzethemindashboardsbasedonitsproprietaryIn
2、-Chiptechnology.”-GartnerMagicQuadrant
3、TalkingdataHOWITALLSTARTED
4、TalkingdataWHATDOFIVEDATAGEEKSTUDENTSDREAMABOUT?
5、TalkingdataWELL,BELIEVINGTHEY’REBADASS…THEY’REDREAMINGOF…
6、TalkingdataBEER&CHIPS
7、TalkingdataINORDERTOUNDERSTANDIN-CHIPANALYTICSDataBeerLET’SASSUMETHAT:
8、TalkingdataMEMORYHIERARCHY
9、INMODERNCPUSCPURAMRAMRAMDISKL1CacheL2CacheL3CacheMainMemoryDiskCapacity:Capacity:Capacity:Capacity:Capacity:64KB-256KB-1MB6MB-20MBGBs-TBsUnlimited128KBLatency:Latency:Latency:Latency:Latency:3Cycles10Cycles35Cycles150-450Cycles1MCycles
10、TalkingdataL2CacheL3CacheMainMemoryx3x50Uptox10x100SO,WH
11、YSlowdownSlowdownSlowdownSHOULDWEEVENCARE?SlowdownwhenfetchingnewdatatotheCPU
12、TalkingdataMEMORYBANDWIDTHIfdataequalsbeerthendatastorageunitsequalalltheplacesbeeriskept!x1L1cacheHomefridgeDistanceImmediateCustomerx10L2/l3cacheShopDistanceBicycleCustomerx50SupermarketDistanceCarCustomerRamDisk
13、BreweryDistanceAirplaneCustomer
14、TalkingdataTHERESHOULDHAVEBEENASLIDEHERE..(it’sthebeer’sfault…)
15、TalkingdataHowdoesSisenseovercomethememorybottleneck?HOWDOESSISENSEOVERCOMETHEMEMORYBOTTLENECK?StorealldataontheDiskOnlyUseRAMWhenaQueryRunsLoadOnlytheRelevantColumnsinRAM
16、TalkingdataVECTORIZATION
17、JITLLVM&SIMDVECTORIZATION&CACHEAWARENESSFirstResultColumn1Column2Column3Column4intoRAMVector1004K100100100100100(Values)4K4K4K4K4K(Values)(Values)(Values)(Values)(Values)OPL1CacheSIMDREGISTEROPApplyOperationOn4/8DataElements100OPSimultaneously4K(Values)100ResultVector4K(Values)1004K(Values)P
18、ushBackToRAM
19、TalkingdataJITLLVMCOMPILATIONWITHSIMDSUPPORT“SIMD”(SingleInstruction,MultipleData)istheprocessofrewritingaloopsothatinsteadofprocessingasingleelementofanarrayNtimes,itprocesses(say)4elementsofthearraysimultaneouslyN/4times.intf{inta,in