物联网和人工智能领域内置芯片分析的意外之旅

1、Future-proofing BI:an unexpected journey to leverage In-Chip analytics in IoT and AIAni ManianHead of Product StrategySIMPLIFYING Business Analytics for COMPLEX Data- Gartner Magic Quadrant| Talking data“The key strength of Sisense is the platforms capability to easily handle and manage large and di

2、verse datasets, and analyze them in dashboards based on its proprietary In-Chip technology.”HOW IT ALL STARTED| Talking dataWHAT DO FIVE DATA GEEK STUDENTS DREAM ABOUT?| Talking dataWELL, BELIEVING THEYRE BADASS THEYRE DREAMING OF| Talking dataBEER & CHIPS| Talking dataIN ORDER TO UNDERSTANDIN-CHIPA

3、NALYTICS| Talking dataLETS ASSUME THAT:MEMORY HIERARCHY IN MODERN CPUSCPURAMDISK L1 CacheCapacity:64KB- 128KBLatency:3 Cycles L2 Cache L3 Cache Main Memory DiskCapacity:Capacity:Capacity:GBs-TBsCapacity:256KB-1MB6MB-20MBUnlimited Latency:10 Cycles Latency:35 Cycles Latency:150-450 Cycles Latency:1M

4、Cycles| Talking dataL2 CacheL3 Cachex10Main Memoryx3 Slowdownx50Up toSO,WHYx100 Slowdown SlowdownSHOULDWE EVEN CARE?Slowdown when fetching new data to the CPU| Talking dataMEMORY BANDWIDTHIf data equals beer then data storage units equal all the places beer is kept!x1DistanceL1 cacheHome fridgeImmed

5、iateCustomerx10DistanceL2 / l3 cacheShopBicycleCustomerx50DistanceSupermarketCarCustomerRamDiskBreweryDistanceAirplaneCustomer| Talking dataTHERE SHOULD HAVE BEEN A SLIDE HERE.(its the beers fault)| Talking dataHow does Sisense overcome the memory bottleneck?HOW DOES SISENSE OVERCOME THE MEMORY BOTT

6、LENECK?Store all data on the DiskOnly Use RAM When a Query RunsLoad Only the Relevant Columns in RAM| Talking dataVECTORIZATIONJIT LLVM & SIMDVECTORIZATION & CACHE AWARENESSFirst into RAMColumn 2Column 4SIMD REGISTERApply Operation On 4/8 Data Elements SimultaneouslyOOPResult Vector1004K(Values)Push

7、 Back To RAM| Talking dataL1 Cache1004K(Values)1004K(Values)1004K(Values)Column31004K(Values)P1004K(Values)1004K(Values)Result Vector1004K(Values)Column 11004K(Values)OPJIT LLVM COMPILATION WITH SIMD SUPPORT“SIMD” (Single Instruction, Multiple Data) is the process of rewriting a loop so that instead

8、 of processing a single element of an array N times, it processes (say) 4 elements of the array simultaneously N/4 f int a, int b)if a ”10” OR f4 = “0” OR f4 = “1”)Field Vector = ValueMask Vector = True / Falsef2 Vectorf3 Vectorf4 Vectorf1 Vector“Customer 1”“Customer 2”“Beer 1”“Beer 2”10=0

9、/=1OROROROR“1”/”2”/”3”Mask Vector&Mask VectorMask VectorMask Vector| Talking dataL1 CacheNEXT: PERFORMANCE TUNING FOR MANY USERSADD HARDWAREADD INSTANCESOPTIMIZE DATA MODEL HOW CAN YOU DELAY USING THESE OPTIONS?| Talking dataPROBLEM: THE WAITING LINE TO QUERY DATAThe queue means a user wait is exten

10、ded by each user in front of themUSERSSECONDS| Talking dataQUERYS BUILDING BLOCKS: THE INSTRUCTION SETS| Talking dataCROWD SPEED: MACHINE LEARNING ARCHITECTUREBreak each query into partsQUERY EXECUTION SPEEDStore each query part and learnBuild new queries with matching parts to boost performance| Ta

11、lking dataRE-USE REPEATING INSTRUCTION SETS ACROSS QUERIES#3WHO WERE THE TOP SALES REPS EACH MONTH?#2WHAT WERE THE MONTHLY SALES?#1HOW MANY UNITS DID WE SELL?Similar but non-identical queriesNew QueryAlready calculated units soldAlready calculated units sold & Monthly breakdown of units| Talking dat

12、aMACHINE LEARNING BIWith Machine Learning BI, analytics get faster even when queries are not identical. The more questions you throw at it - the more efficient it gets!More users = more queries = faster resultsNo matchMatch found| Talking dataIN-CHIP = POWER + MACHINE LEARNINGLeverage the unique in-

13、chip cache memory to perform faster than in-memoryWithout the limitation of having to load the entire model into RAMIn-Chip recognizes the CPU specs and applies its unique code to organize the query data in the CPUWhen needed again, that piece of data exists in the CPU cache, which is much faster th

14、an RAMIn-Chip machine-learns to fetch the associated compressed result sets in advanceSub-query results pre-loaded into L1 cache as compressed dataDecompressed images of that same data can be moved to the larger, but slower, L2 and L3 cachesDecompression operations (read from and write to cache) are

15、 extremely fast| Talking dataIN-CHIP TECHNOLOGYThe best engine beer can buyIn memory columnar execution modeColumnar storage CACHE aware query kernelFull 64BIT support CACHE aware decompressionLLVM based compiler with SIMD support Instruction recycling & learning algorithms| Talking dataEMPOWERING G

16、ROWTH, ANYWHERE, EVERYWHERE, ON AFFORDABLEHWIN-CHIP BI BENCHMARKSkylakeHaswellBENCHMARK SETTINGS160140Dataset:120M rows 28GB120100300%faster808 Analytical queries X 50 cycles6040Aggregations Grouping Top RankingLarge intermediate results200Test 1: No concurrencyTest 2: Concurrency = 2Test 3: Concurr

17、ency = Max| Talking dataSPEED! STRATA AWARDAnalyzing 10TB of data in 10 secondsOn a single node on a standard Dell Server| Talking dataREVOLUTION: SCALE-OUT VS IN-CHIPIn-ChipArchitectureScale-OutData Scientists, IT, DevelopersETL, Batch Reports, Machine LearningJAVA, R, C, SQLLong BigBusiness Users

18、Ad-Hoc AnalyticsInteractive Dashboards, SQLShortSmallUsers Use Cases InterfaceTime to ImplementAvailable ResourcesBig Data InfrastructureAgile Big Data AnalyticsOutcome| Talking dataSOWHAT IS IT GOOD FOR?| Talking dataFROM COMPLEXITYTO SIMPLICITY| Talking dataTECHNOLOGY HAS NO MEANING IF IT HAS NO I

19、MPACT ON HUMAN LIFE“If a tree falls in a forest and no one is around to hear it, does it make a sound?”| Talking dataOVERHYPE OF BUZZWORDS| Talking dataTHE PERSONAL, INTELLIGENT AND CONTEXTUAL WEB| Talking dataTHE INTERNET OF ME| Talking dataTRANSFORMATION OF BIG DATA ANALYTICS FOR IOM| Talking data

20、WE ARE ALL UNIQUE| Talking dataLET THERE BE LIGHTS| Talking dataTHE NEXT REALM OF BUSINESS ANALYTICSAnalyzing data no longer requires being anchored to a screenSisense Everywhere devices broadcast business KPIs to all the senses Making consumption of insights immediate and simple.| Talking dataHOW I

21、T ALL STARTED| Talking dataThe relationship between business professionals and their KPIs(We asked hundreds of business professionals how they interact with their data and KPIs)ALMOST HALF OF ALL RESPONDENTS CHECK KPISDAILYHow often do you check the status of your KPIs?| Talking data83% OF RESPONDEN

22、TS USE OR WANT TO USE COLOR CODINGDo you use color-coding in the way you display data?| Talking dataVISUAL ALERTS ARE THE MOST EFFECTIVE, ACCORDING TO MORE THANHALF OF RESPONDENTSWhich type of alert is best in driving you to action?| Talking dataACCORDING TO RESPONDENTS THE FUTURE OF BI CONSUMPTIONI

23、SEVERYWHEREHow would you like to consume data in the future?| Talking dataBI EVERYWHERERevolutionizing The Way Business Users Consume Data| Talking dataIMAGINE A SISENSE WORLD| Talking dataImagine youre driving to work and can ask your voice- operated BI assistant: What is my sales target for today?

24、IMAGINE A SISENSE WORLD| Talking dataImagine being able to focus your entire team on improving customer satisfaction just by having them glance at the Sisense IoT bulb,green means on target and red means take action.IMAGINE A SISENSE WORLD| Talking dataImagine stepping into a conference room fora qu

25、arterly business review and experiencing your data insights hovering around you.SISENSE BRINGS IMAGINATION TO LIFESisense-Enabled is a new line of devices that present data unlike any dashboard environmentSISENSE LAMPSISENSE ENABLED ECHO| Talking dataREDEFINING HOW WE INTERACT WITH DATA“When I see that bulb change, I get a real sense of satisfaction. Its provided a direct way for us to see how data is changing. The bulb gives me peace of mind because I can see a light change rather than monitoring a screen.”RESPOND TO CHANGES IN REAL-TIME“I