A Relational Algebra Processor - PowerPoint Presentation

Slide1

A Relational Algebra Processor

6.375 Final Project

Ming Liu,

Shuotao

XuSlide2

Motivation

Today’s Database Management Systems (DBMS): software running on a standard operating system on a general purpose CPU

DBMS frequently used in analytics and scientific computing, but bottlenecked by:

Processor speed, software overhead, latency & bandwidthProposal: FPGA Based Relational Algebra Processor

Host PC

(DBMS)

FPGARelational Algebra Processor

Physical Storage

2Slide3

Background|Relational

Algebra (RA)

Many database

queries are fundamentally decomposable to five basic RA operatorsAlthough SQL is capable of much more

Operator

Functions

SelectionFilter rows based on a Boolean condition

ProjectionEliminate selected attributes (columns) of a table; remove duplicated results

Cartesian Product

Combine several tables with unique attributesUnion

Combine several tables with the same attributes

Difference

Select rows of several tables where the rows do not match

Design dedicated processors on the FPGA for each operator

3Slide4

Project Goal

Design and implement an in-memory relational algebra processor on the FPGA

Explore the types of queries that can benefit from FPGA acceleration

Secondary: Outperform SQLite!Some assumptions:32-bit wide table entries

Tables fit in memoryMax number of columns is 32

Read only4Slide5

Microarchitecture | Host Software

5

FPGASlide6

Microarchitecture |

Top-Level

RAProcessor

Host PC

(C++ functions)

RA Processor

DRAM

PCIe

Host PC

(DBMS)

RA Processor

Physical Storage

6Slide7

Microarchitecture | Row

Marshaller

7

Exposes a simple interface for operators to access tables in DRAM

Address translation, burst aggregation, truncation & alignment

Multiplexes requestsTable values sent/received as 32-bit burstsSlide8

Microarchitecture | Selection

8

Filters rows based on predicates (e.g. age < 40)

16 predicate evaluators

Internally comparators

A tree of gates to qualify the predicatesMax: 4 ORs of 4 ANDsSlide9

Microarchitecture | Projection

9

Select columns of a table

Column mask one-hot encoded

Do not need to buffer row; operate directly on data burstsSlide10

Microarchitecture | Binary Operators

10

Cartesian

Product, Union, Difference and

Deduplication

Nested loop implementationSlide11

Microarchitecture|

Inter-operator

Bypassing11

Operators enabled concurrently; data passed between operators

No intermediate storage Conditions:

A singly link of unary operatorsEach operator has a single target outputNo structural

hazardSoftware reorders and schedules the RA commandsData source/destination encoded in commandSlide12

Microarchitecture|

Inter-operator

Bypassing12

Multiple 32-bit wide output FIFOs to other operatorsSlide13

Implementation Evaluation

13

Timing

Maximum Frequency: 55.786MHz

Critical Path: Row

Marshaller muxArea

Slice Registers: 50%LUTs: 85%BRAM/FIFOs: 47%

Modules

Slice Registers

LUTs

BRAM/FIFOs

TOTAL

34649

(50%)

59328 (85%)71 (47%)

Row

Mashaller

2804

6627

0

Controller

4570

6277

29

Selection

3137

19633

0

Projection

739

654

0

Cartesian

Product

1935

1478

0

Union

1939

1983

0

Difference

1875

1949

0

Deduplication

1822

1970

0Slide14

Performance Benchmark |

Setup

14

SQLite

Internal

SQLite timer to report execution time of the queryThinkpad T430, Core i7-3520M @ 2.90Ghz, 1x8GB DDR3-1600

RA ProcessorPerformance counters: cycles from start to

ack of an operator

Table

Relational Algebra Query

SQL Query

1 table

100k

x 30

SELECT,starLong,tableOut, mass,>,80000,AND,pos_x,>,10, OR,pos_x,<,pos_z, OR,col12,>,col14, AND,col20,<,col21SELECT * FROM starLongWHERE mass > 80000 AND pos_x > 10 OR pos_x < pos_z OR col12 > col14 AND col20 < col21;

1 table

100k x

30

PROJECT,starLong,tableOut

,

pos_x,col19,col25,col29

SELECT pos_x,col19

, col25,

col29 FROM

starLong

;

2 tables

1k

x 30

UNION,starMed1,starMed2,starUnion

SELECT * FROM starMed1

UNION

SELECT * FROM starMed2;

2 tables

1k

x 30

XPROD,starMed1,starMed2,starXprod

RENAME,starXprod,0,iOrder0,1,mass0,8,phi0

SELECT,starXprod,starFiltered

, iOrder0,=,

iOrder

, AND,phi0

,>,1

,

AND,mass0,>,mass

PROJECT,starFiltered,starOut,mass0

SELECT

s1.mass

FROM starMed1

s1

, starMed2

s2

WHERE s1.vx

> s2.vx

AND s1.phi

> 1

AND s1.mass

>

s2.mass;Slide15

Performance Benchmark |

Results

15

Limitation: Memory Bandwidth: 200MB/s

vs

12.8GB/sSlide16

Performance Benchmark |

Results

16

Select operator

most competitive with SQLite

What happens with

more predicates?Slide17

Improvements

17

Increasing data burst width

32-bit to 256-bit:

potential 8x speedup

Area/critical path increaseMaximizing memory bandwidth

Additional row buffers to buffer data from DDR2 Memory

Larger, faster DRAM; Higher clock speedSlide18

Conclusion & Future Work

18

Complex filtering operations performs well on the FPGA

Better than SQLite with sufficient memory bandwidthData intensive operators do not perform wellFuture opportunities:

An accelerator alongside SQLiteIntegration with HDD/SSD controller