query processing on smart ssds - university of wisconsin

Query Processing on Smart SSDs

Opportunities and Challenges Jaeyoung Do1, Yang Seok Ki2, Jignesh M. Patel1, Chanik Park2, Kwanghyun Park1, David J. DeWitt3 1 University of Wisconsin – Madison 2 Samsung Semiconductor Inc. 3 Microsoft Jim Gray Systems Lab

1

Contact: [email protected]

The opinions expressed here are those of the authors, and not necessarily of the organizations to which the authors belong.

Motivation 2

CPU

DRAM

caches

Magnetic Hard Disk Drives

CPU

NVRAM (e.g. SSDs)

Bus

NVRAM (e.g. SSDs)

DRAM

Inside a Modern SSD

•  Embedded processors (low power) CPU

• General-purpose memory Memory

• High I/O bandwidth inside the flash chips I/O

3

SSD Controller

Embedded Processor

DRAM Controller

Flash Memory Controller



SRAM

Host Interface Controller

DRAM

F F

F F

F F

Flash SSD Device

4

1

10

100

2007 2008 2009 2010 2011 2012 2013 2014 2015

Ban

dwid

th R

elat

ive

to

the

I/O In

terf

ace

Spee

d

Year

I/O Interface

Inside the SSD

5

•  Push code through the “straw,” thereby drinking smaller amounts through the straw Philosophically similar to what Netezza and Exadata do today, but using “commodity” resources inside the SSD

The Opportunity

•  Higher performance •  Lower energy consumption (Joules/query) •  Lower cost when building balanced systems •  Simpler appliances

Potential Advantages

•  Programming tools for the SSD are primitive •  Not enough “spare” processing power •  Some hardware architectural limitations •  Deeper DBMS implications: query optimization,

buffer pool caching, updates, transaction mgmt., etc.

Current Challenges

6

Host I/O Interface

I/O Interface

Aggregate

Scan

Host I/O Interface

I/O Interface

Aggregate

Pre- Aggregate

Scan

Implementation Details

•  New Smart SSD APIs to run database operations inside the SSD

•  Open(), Get(), Close(), SendToHost()

Samsung Smart SSD

•  Modified the query processor to hard-code selected Smart SSD plans

•  Implement Smart SSD executors

SQL Server

7

Communication Flow – e.g. Scan 8

Scanning … Get(): Are there results?

NO

Get(): Are there results? YES, results

5μsec

Get(): Are there results

FINISHED Scan is done

Open(): Metadata

OK

SQL Server Smart SSD

Close() OK

9

Data page

A tuple

A value of a column in a tuple

Embedded Processor

DRAM Controller




SRAM Host

Interface Controller

DRAM

F F

F F

F F

Inside the Flash SSD

Data storage

If qualified

Copy projected columns’ values to output buffer

int char int vchar

Data Flow inside the SSD (NSM format)

Storage Formats • NSM layout

• PAX layout

10

Data Page

Tuple int char date int char date

Column

11

Embedded Processor

DRAM Controller




SRAM Host

Interface Controller

DRAM

F F

F F

F F

Inside the Flash SSD

Data storage

Data page

A column

A value of a column in a tuple

If qualified

32 Bytes LDM_CMP function

Copy projected columns’ values to output buffer

int

char

int int int

char char char

Data Flow inside the SSD with the PAX layout

System Configuration 12

SQL server (Host machine) Flash SSD HDD 64-bit Windows 7

32GB DRAM (24GB dedicated to the DBMS)

2.66 GHz Dual Quad Core Intel Xeon5430 32KB L1 cache, 6MB L2 cache

Two 7.5 RPM SATA HDDs one for the OS and one for the Log

LSI four-port SATA/SAS 6Gbps HBA (host bus adapter)

Samsung 400GB SAS SSD Two ARM Cores

16KB SRAM for each core

256KB DRAM for each core

1MB DRAM output buffer

HP 146GB 10K RPM SAS HDD

Upper Bound on Expected Performance Gains

0

500

1000

1500

SAS HDD SAS SSD Smart SSD (Internal)

80

550

1560

Max

imum

Seq

uent

ial R

ead

(MB

/s)

13

20X

Speed of a Simple Sequential Scan Program

Experimental Setup: Synthetic Data 14

Table # Columns

Table Size (GB)

# tuples (millions)

# tuples/page

Synthetic 4 4 10 400 323

Synthetic16 16 30 400 109

Synthetic64 64 110 400 29

Each table has a number of integer-typed columns

Experimental Setup: TPC-H 15

Lineitem ( L_ORDERKEY bigint, L_PARTKEY int, L_SUPPKEY int, L_LINENUMBER int, L_QUANTITY bigint, L_EXTENDEDPRICE bigint, L_DISCOUNT bigint, L_TAX bigint, L_SHIPDATE int, L_COMMITDATE int, L_RECEIPTDATE int, L_RETURNFLAG char(1), L_LINESTATUS char(1), L_SHIPINSTRUCT char(25), L_SHIPMODE char(10), L_COMMENT char(47), )

Modified TPC-H Lineitem Table, 600M tuples, 94GB

10% Selection Query

16

1

10

100

1000

10 100 1000

Tota

l Ene

rgy

(KJ)

, lo

g sc

ale

Elapsed Time (seconds), log scale

15.5

X

16.0X

SELECT COLUMN_2 FROM Synthetic64 WHERE COLUMN_1 < [VALUE] Smart SSD (PAX)

Smart SSD (NSM)

SAS SSD

SAS HDD

10% Selection Query

17

10

100

10 100

Tota

l Ene

rgy

(KJ)

, lo

g sc

ale


2.4X

2.3X


Smart SSD (NSM)

SAS SSD

SAS HDD

10% Selection Query

18

0

5

10

15

0 500 1000 1500

I/O S

ubsy

stem

Ene

rgy

(KJ)

Elapsed Time (seconds)

SELECT COLUMN_2 FROM Synthetic64 WHERE COLUMN_1 < [VALUE]

23.0

X

16.0X

Smart SSD (PAX)

Smart SSD (NSM)

SAS SSD

SAS HDD

10% Selection Query

19

0

0.5

1

0 50 100 150 200 250

I/O S

ubsy

stem

Ene

rgy

(KJ)



Smart SSD (NSM)

SAS SSD

SAS HDD

1.8X

2.3X

Effect of the # of Tuples on a

0

50

100

150

200

Synthetic4 Synthetic16 Synthetic64

Tim

e (s

econ

ds)

Target Table

SAS SSD Smart SSD(PAX)

SELECT COLUMN_2, FROM [Synthetic_table] WHERE COLUMN_1 < [VALUE]

Synthetic4: (323 tuples/page) Synthetic16: (109 tuples/page) Synthetic64: (29 tuples/page)

Smart SSD outperforms the regular SSD Smart SSD

underperforms compared to

the regular SSD

10% Aggregation Query

21

1

10

100

1000

10 100 1000

Tota

l Ene

rgy

(KJ)

, lo

g sc

ale


16.4

X

16.9X

SELECT AVG(COLUMN_2) FROM Synthetic64 WHERE COLUMN_1 < [VALUE] Smart SSD (PAX)

Smart SSD (NSM)

SAS SSD

SAS HDD


22

10

100

10 100

Tota

l Ene

rgy

(KJ)

, lo

g sc

ale


2.6X

2.4X


Smart SSD (NSM)

SAS SSD

SAS HDD


23

0

5

10

15

0 500 1000 1500

I/O S

ubsy

stem

Ene

rgy

(KJ)


SELECT AVG(COLUMN_2) FROM Synthetic64 WHERE COLUMN_1 < [VALUE]

20.9

X

16.9X

Smart SSD (PAX)

Smart SSD (NSM)

SAS SSD

SAS HDD


24

0

0.5

1

0 50 100 150 200 250

I/O S

ubsy

stem

Ene

rgy

(KJ)



Smart SSD (NSM)

SAS SSD

SAS HDD

1.8X

2.4X

TPC-H Query 6

25

10

100

10 100 1000

Tota

l Ene

rgy

(KJ)

, lo

g sc

ale


SELECT SUM(EXTENDEDPRICE*DISCOUNT) FROM LINEITEM WHERE SHIPDATE>=1994-‐01-‐01, SHIPDATE<1995-‐01-‐01, DISCOUNT > 0.05, DISCOUNT < 0.07, QUANTITY < 24

11.8

X

11.5X

Smart SSD (PAX)

Smart SSD (NSM)

SAS SSD

SAS HDD

TPC-H Query 6

26

10

100

10 100

Tota

l Ene

rgy

(KJ)

, lo

g sc

ale


1.9X

1.7X


Smart SSD (PAX)

Smart SSD (NSM)

SAS SSD

SAS HDD

TPC-H Query 6

27

0

5

10

15

20

0 200 400 600 800 1000 1200

I/O S

ubsy

stem

Ene

rgy

(KJ)


14.1

X

11.5X


Smart SSD (PAX)

Smart SSD (NSM)

SAS SSD

SAS HDD

TPC-H Query 6

28

0

0.5

1

1.5

2

0 50 100 150 200

I/O S

ubsy

stem

Ene

rgy

(KJ)



Smart SSD (PAX)

Smart SSD (NSM)

SAS SSD

SAS HDD

1.4X

1.7X

Conclusions and Directions for Future Work •  Computing and storage boundaries are no longer as

rigid as they have been in the past •  “Commodity” computing in storage devices is here •  Communication buses evolve slower than other parts

of the system

Big Picture

•  Push code to data rather than pull data to the (main) processor

•  Make general-purpose processing a commodity pre-built “into” the device

Opportunity

•  Need better programming tools •  Various SSD hardware architectural changes needed

(but appear to be feasible, cheaply) •  DBMS internals also need significant changes

Challenges

29

Co-dependency

Hardware •  More processing

behind the bus •  Commodity

processing •  Better programming

tools •  Long-term roadmap

Software •  Caching

implications •  Query optimization •  Consistency •  Concurrency •  Storage

optimization

30

•  Significant $ and time investment needed on both sides.

•  Who is willing to take the long-term view and invest?

query processing on smart ssds - university of wisconsin

Documents