云原生數(shù)據(jù)倉庫TPC-H第一背后的Laser引擎大揭秘+-++阿里云-+魏闖先

云原生數(shù)據(jù)倉庫TPC-H第一背后的Laser引擎大揭秘魏闖先阿里云數(shù)據(jù)庫產(chǎn)品事業(yè)部內(nèi)容ADB

PG和Laser計(jì)算引擎介紹Laser計(jì)算引擎技術(shù)未來展望PART01ADB

PG和Laser計(jì)算引擎介紹ADB

PG架構(gòu)MasterInterconnect

NetworkExecution

&StorageNodeParserLogical

PlanOptimizerPhy.

Plan

Native

ExecEngineVectorizedExecutionEnginePlanTransformFusionScanBatchDataManagerLLVM

CodegenHardwareFriendly

ImplementationEfficientMotionMemoryContextSortAggJoinOthersSortJitterAggJitterJoinJitterScanJitterScanOthersStorageRow

Oriented

TableColumn

Oriented

TableForeign

Table

OSSExecutionExecution

&StorageNodeExecution

&StorageNodeMasterMaster為什么做Laser計(jì)算引擎AggTupleProjectexpr=(c3+30)*50TupleFilterexpr=c1+c2>1000TupleScanC330+50*call()call()call()call()過多函數(shù)調(diào)用C1C2+1000>call()call()call()call()過多函數(shù)調(diào)用call()call()例子：SELECTSUM((c3+30)*50)ASBonusFROMstore_salesWHEREc1+c2>1000PostgreSQL計(jì)算引擎問題：重存儲TP場景執(zhí)行樹遞歸調(diào)用逐行處理數(shù)據(jù),無列存優(yōu)化序列化開銷高，內(nèi)存多次拷貝單分區(qū)單進(jìn)程執(zhí)行主流OLAP業(yè)界方案Vectorization做法類似Volcano模型，采用批量模式執(zhí)行PULL模式（自上而下）優(yōu)點(diǎn)減少cache

miss適合SIMD加速缺點(diǎn)實(shí)現(xiàn)邏輯復(fù)雜不適合OLTP場景Batch緩存開銷典型代表DuckDBClickHouseCompilation做法使用codegen將所有算子編譯成一個(gè)函數(shù)PUSH模式（自下而上）優(yōu)點(diǎn)無需緩存，可寄存器訪問適合計(jì)算復(fù)雜性加速缺點(diǎn)PUSH模式控制復(fù)雜Cache

miss率高典型代表SparkPelotonVS業(yè)界主流計(jì)算引擎對比產(chǎn)品ClickHousePG

11SparkADB

PG執(zhí)行模型向量引擎火山模型編譯執(zhí)行向量引擎內(nèi)存模型列存行存行存行/列存即時(shí)編譯表達(dá)式級LLVM表達(dá)式級LLVMStage

JAVA算子級LLVM硬件加速不支持不支持FPGA支持GPUPART02Laser計(jì)算引擎技術(shù)Laser計(jì)算引擎向量計(jì)算引擎行列式內(nèi)存模型JIT加速SIMD指令加速FUSION

SCANTable

CustomerTable

OrdersTable

LineitemBATCHJITSCANData

BatchDataArrayNullBitmapOffsetArrayc_mktsegment='FURNITURE'o_orderdate<'1992-05-01'l_shipdate>'1992-05-01'JoinCondition:c_custkey=

o_custkeyJoinCondition:l_orderkey=o_orderkeyGroupByKey:l_orderkey,o_orderdate,o_shippriorityBATCHJITSCANBATCHJITSCANBATCHJITHash

JoinBATCHJITHash

AggBATCHJITHash

Join向量化執(zhí)行減少虛函數(shù)調(diào)用SIMD指令優(yōu)化多個(gè)函數(shù)合并Cache

line對齊Zerocopy/內(nèi)存池化intsum(x,y){

returnx+y;}volcano_execute(){for(i=0;i<N;i++)

sum(result,data[i]);}int[]sum(x[],y[]){

for(i=0;i<2048;i++)ret[i]=x[i]+y[i];}laser_execute(){for(i=0;i<N/2048;i++)

sum(result,data

+

i*

2048)}Q1提升120%Q18提升190%SIMD指令加速SELECTsum(price*ratio)asrevenueFROMlineitem1.8億行l(wèi)ineitem表測試性能提升4X

Just-in-Time編譯優(yōu)化IRMachine

code//樣例SQL

selectcount(*)fromtablewherea>

10andb<

5;

//PostgreSQL解釋執(zhí)行方案：多次函數(shù)調(diào)用

result=

Int8AndOp(Int32GT(a,

10),

Int32LT(b,

5));

//使用LLVMcodegen生成最小化底層代碼

%res1=icmpugti32%a,

10;

%res2=icmpulti32%b,

5;%res=andi8%res1,

%res2;相對于解釋執(zhí)行，JIT性能提升20~30%FusionScanSELECT

A,

C*D

FROM

TABLE

WHERE

A

>

3and

B<6;63100.50.20.6100200300Column

A

只讀取A、B列數(shù)據(jù)解析Block數(shù)據(jù)格式過濾計(jì)算A列數(shù)據(jù)寫入Batch610Batch13RowIdArrayColumn

C61020180讀取C和D列數(shù)據(jù)解析Block數(shù)據(jù)格式跳過無效數(shù)據(jù)計(jì)算表達(dá)式(C*D)并寫入Batch返回Batch和有效行數(shù)Column

D384LLVMIR

A>3andB<6Column

BLLVMIR

C*DScan算子性能提升3X算子實(shí)現(xiàn)-HashJoin數(shù)據(jù)Hash

EntryHash

Entry

1Hash

Entry

2Hash

Entry

3

Hash

Entry

N…開放鏈表Hash

ValNext

RowRowIdRowIdRowIdRowIdHash

ValNext

RowHash

ValNext

RowHash

ValNext

RowHash

ValNext

RowHash

ValNext

RowHash

ValNext

Row…Hash表信息Hash

Entry

3

RowId獲取行號生成結(jié)果比Hash值比Hash

key結(jié)束RowId=-1FalseFalseHash值/Entry插件集成開啟LaserLaserRunHookLaserEndHookLaserStartHookCreateExtensionLaser_PG_initPostgres內(nèi)核HookParserPlannerOptimizerOnLaserNativeEngine插件機(jī)制YNTPC-H結(jié)果相比第二名微軟SQLServ