LEVERAGING DSLS (ALMOST) FOR FREE @alvinkcheungmetalift.uwplse.org/strangeloop17.pdf · [17] K....

LEVERAGING DSLS (ALMOST) FOR FREE

@alvinkcheung

uwdb.iouwplse.org

Time

Analytics

ImageProcessing

Hardware

ParallelComputing

Rotate Blur

Join

CachingPipelining

Hashimplementation

Embedded algorithmic detailsArchitecture-specific optimizations

HashPartitioning

?YourDSLhere

Annuals of Code Rewrites

• Overhaul of the Mozilla Gecko layout engine (2 years),and now to Servo

• Rewrite of INGRES database into PostGRES (3 years)

• Twitter search engine rewrite (2 years)

All of these were major engineering efforts!

Vision:Perform code rewrites automatically

Current focus: Performance-critical code fragments

for ($i = 0; $i < $l1.size(); ++$i) {$l2.append($l1[$i] + $c);

}$l2 = $l1.map(e -> e + $c);

List getUsersWithRoles () {List users = this.userDao.getUsers(); List roles = this.roleDao.getRoles(); List results = new ArrayList();for (User u : users) {

for (Role r : roles) {if (u.roleId == r.id)

results.add(u); }}

return results; }

?

Syntax-Driven Compilation

[Radoi et al, OOPSLA 14]

+ E:localize-fold-accssesdocs . f o l d (m) ( {

c a s e ( v34 , ( i , v8 ) ) =>val v4 = v8 . s p l i t ( " " ) . groupBy ( ID ) . __2 ;val v5 = v4 . z i p ( v34 ) . map ( {

c a s e ( v2 , ( v3 , v7 ) ) =>v7 + v3 . s i z e

} ) ;v34 ++ v5

} )

+ E:map-vertical-fissiondocs . f o l d (m) ( {

c a s e ( v34 , ( i , v8 ) ) =>val v4 = v8 . s p l i t ( " " ) . groupBy ( ID ) . __2 ;val v5 = v4 . map ( {

c a s e ( v2 , v3 ) => v3 . s i z e} ) . z i p ( v34 ) . map ( {

c a s e ( v2 , ( v9 , v7 ) ) => v7 + v9} ) ;

v34 ++ v5} )

+ E:identify-map-monoid-plusdocs . f o l d (m) ( {

c a s e ( v34 , ( i , v8 ) ) =>( v8 . s p l i t ( " " ) . groupBy ( ID ) . __2 ) . map ( {

c a s e ( v2 , v3 ) => v3 . s i z e} ) | + | v34

} )

+ E:pull-map-from-fold-by-subexpression-extractdocs . map ( {

c a s e ( i , v8 ) =>( v8 . s p l i t ( " " ) . groupBy ( ID ) . __2 ) . map ( {

c a s e ( v2 , v3 ) => v3 . s i z e} )

} ) . f o l d (m) ( {c a s e ( v34 , ( i , v10 ) ) =>

v10 | + | v34} )

+ E:map-horizontal-fissiondocs . map ( {

c a s e ( i , v8 ) =>v8 . s p l i t ( " " ) . groupBy ( ID ) . __2

} ) . map ( {c a s e ( i , v11 ) =>

v11 . map ( {c a s e ( v2 , v3 ) => v3 . s i z e

} )} ) . f o l d (m) ( {

c a s e ( v34 , ( i , v10 ) ) =>v10 | + | v34

} )

+ E:swap-map-with-foldval v11 = docs . map ( {


} ) . f o l d ( "ZERO�TOKEN" ) ( {c a s e ( v12 , ( i , v13 ) ) =>

v12 | + | v13} ) ;


} )

+ E:map-horizontal-fissionval v11 = docs . map ( {

c a s e ( i , v8 ) =>v8 . s p l i t ( " " ) . groupBy ( ID )

} ) . map ( {c a s e ( i , v14 ) => v14 . __2

} ) . f o l d ( "ZERO�TOKEN" ) ( {

c a s e ( v12 , ( i , v13 ) ) =>v12 | + | v13

} ) ;v11 . map ( {

c a s e ( v2 , v3 ) =>v3 . s i z e

} )


c a s e ( i , v8 ) =>v8 . s p l i t ( " " )

} ) . map ( {c a s e ( i , v15 ) =>

v15 . groupBy ( ID )} ) . map ( {

c a s e ( i , v14 ) => v14 . __2} ) . f o l d ( Map ( ) ) ( {

c a s e ( v12 , ( i , v13 ) ) =>v12 | + | v13

} ) ;v11 . map ( {

c a s e ( v2 , v3 ) =>v3 . s i z e

} )


c a s e ( i , v8 ) =>v8 . s p l i t ( " " )

} ) . map ( {c a s e ( i , v15 ) =>

v15 . groupBy ( ID )} ) . f o l d ( Map ( ) ) ( {

c a s e ( v16 , ( i , v17 ) ) =>v16 | + | v17

} ) ;( v14 . __2 ) . map ( {

c a s e ( v2 , v3 ) =>v3 . s i z e

} )


c a s e ( i , v8 ) =>v8 . s p l i t ( " " )

} ) . f o l d ( Map ( ) ) ( {c a s e ( v18 , ( i , v19 ) ) =>

v18 | + | v19} ) ;

val v14 = v15 . groupBy ( ID ) ;( v14 . __2 ) . map ( {

c a s e ( v2 , v3 ) =>v3 . s i z e

} )

+ E:flatMapval v15 = docs . f l a t M a p ( {

c a s e ( i , v8 ) =>v8 . s p l i t ( " " )

} ) ;val v14 = v15 . groupBy ( ID ) ;( v14 . __2 ) . map ( {

c a s e ( v2 , v3 ) =>v3 . s i z e

} )

+ E:reducebykeydocs . f l a t M a p ( {

c a s e ( i , v8 ) =>v8 . s p l i t ( " " )

} ) . map ( {c a s e ( j , v6 ) =>

( v6 , 1 )} ) . reduceByKey ( {

c a s e ( v2 , v3 ) =>v2 + v3

} )

+ E:localize-fold-accssesdocs . f o l d (m) ( {

c a s e ( v34 , ( i , v8 ) ) =>val v4 = v8 . s p l i t ( " " ) . groupBy ( ID ) . __2 ;val v5 = v4 . z i p ( v34 ) . map ( {

c a s e ( v2 , ( v3 , v7 ) ) =>v7 + v3 . s i z e

} ) ;v34 ++ v5

} )

+ E:map-vertical-fissiondocs . f o l d (m) ( {

c a s e ( v34 , ( i , v8 ) ) =>val v4 = v8 . s p l i t ( " " ) . groupBy ( ID ) . __2 ;val v5 = v4 . map ( {

c a s e ( v2 , v3 ) => v3 . s i z e} ) . z i p ( v34 ) . map ( {

c a s e ( v2 , ( v9 , v7 ) ) => v7 + v9} ) ;

v34 ++ v5} )

+ E:identify-map-monoid-plusdocs . f o l d (m) ( {

c a s e ( v34 , ( i , v8 ) ) =>( v8 . s p l i t ( " " ) . groupBy ( ID ) . __2 ) . map ( {

c a s e ( v2 , v3 ) => v3 . s i z e} ) | + | v34

} )

+ E:pull-map-from-fold-by-subexpression-extractdocs . map ( {

c a s e ( i , v8 ) =>( v8 . s p l i t ( " " ) . groupBy ( ID ) . __2 ) . map ( {

c a s e ( v2 , v3 ) => v3 . s i z e} )

} ) . f o l d (m) ( {c a s e ( v34 , ( i , v10 ) ) =>

v10 | + | v34} )

+ E:map-horizontal-fissiondocs . map ( {


} ) . map ( {c a s e ( i , v11 ) =>


} )} ) . f o l d (m) ( {

c a s e ( v34 , ( i , v10 ) ) =>v10 | + | v34

} )



} ) . f o l d ( "ZERO�TOKEN" ) ( {c a s e ( v12 , ( i , v13 ) ) =>

v12 | + | v13} ) ;


} )


c a s e ( i , v8 ) =>v8 . s p l i t ( " " ) . groupBy ( ID )

} ) . map ( {c a s e ( i , v14 ) => v14 . __2

} ) . f o l d ( "ZERO�TOKEN" ) ( {

c a s e ( v12 , ( i , v13 ) ) =>v12 | + | v13

} ) ;v11 . map ( {

c a s e ( v2 , v3 ) =>v3 . s i z e

} )


c a s e ( i , v8 ) =>v8 . s p l i t ( " " )

} ) . map ( {c a s e ( i , v15 ) =>

v15 . groupBy ( ID )} ) . map ( {

c a s e ( i , v14 ) => v14 . __2} ) . f o l d ( Map ( ) ) ( {

c a s e ( v12 , ( i , v13 ) ) =>v12 | + | v13

} ) ;v11 . map ( {

c a s e ( v2 , v3 ) =>v3 . s i z e

} )


c a s e ( i , v8 ) =>v8 . s p l i t ( " " )

} ) . map ( {c a s e ( i , v15 ) =>

v15 . groupBy ( ID )} ) . f o l d ( Map ( ) ) ( {

c a s e ( v16 , ( i , v17 ) ) =>v16 | + | v17

} ) ;( v14 . __2 ) . map ( {

c a s e ( v2 , v3 ) =>v3 . s i z e

} )


c a s e ( i , v8 ) =>v8 . s p l i t ( " " )

} ) . f o l d ( Map ( ) ) ( {c a s e ( v18 , ( i , v19 ) ) =>

v18 | + | v19} ) ;

val v14 = v15 . groupBy ( ID ) ;( v14 . __2 ) . map ( {

c a s e ( v2 , v3 ) =>v3 . s i z e

} )

+ E:flatMapval v15 = docs . f l a t M a p ( {

c a s e ( i , v8 ) =>v8 . s p l i t ( " " )

} ) ;val v14 = v15 . groupBy ( ID ) ;( v14 . __2 ) . map ( {

c a s e ( v2 , v3 ) =>v3 . s i z e

} )

+ E:reducebykeydocs . f l a t M a p ( {

c a s e ( i , v8 ) =>v8 . s p l i t ( " " )

} ) . map ( {c a s e ( j , v6 ) =>

( v6 , 1 )} ) . reduceByKey ( {

c a s e ( v2 , v3 ) =>v2 + v3

} )

[13] M. H. Hall, S. P. Amarasinghe, B. R. Murphy, S.-W. Liao,and M. S. Lam. Detecting coarse-grain parallelism using aninterprocedural parallelizing compiler. Supercomputing ’95,1995.

[14] R. Joshi, G. Nelson, and K. Randall. Denali: A goal-directedsuperoptimizer. PLDI ’02, pp. 304–314, 2002.

[15] R. A. Kelsey. A correspondence between continuation passingstyle and static single assignment form. IR ’95, pp. 13–22,1995.

[16] Y. Klonatos, A. Nötzli, A. Spielmann, C. Koch, and V. Kuncak.Automatic synthesis of out-of-core algorithms. SIGMOD ’13,pp. 133–144, 2013.

[17] K. Knobe and V. Sarkar. Array SSA form and its use inparallelization. POPL ’98, pp. 107–120, 1998.

[18] R. Lämmel. Google’s MapReduce programming model —revisited. Science of Computer Programming, 70(1):1 – 30,2008.

[19] S.-w. Liao. Parallelizing user-defined and implicit reductionsglobally on multiprocessors. ACSAC’06, pp. 189–202, 2006.

[20] E. Meijer, M. Fokkinga, and R. Paterson. Functional program-ming with bananas, lenses, envelopes and barbed wire. FPCA’91, pp. 124–144, 1991.

[21] C. Nugteren and H. Corporaal. Introducing Bones: a paralleliz-ing source-to-source compiler based on algorithmic skeletons.GPGPU-5, pp. 1–10, 2012.

[22] B. C. Oliveira, A. Moors, and M. Odersky. Type classes asobjects and implicits. OOPSLA ’10, pp. 341–360, 2010.

[23] N. Ramsey. Unparsing expressions with prefix and postfixoperators. Software: Practice and Experience, 28(12):1327–1356, 1998.

[24] C. Ranger, R. Raghuraman, A. Penmetsa, G. Bradski, andC. Kozyrakis. Evaluating MapReduce for multi-core andmultiprocessor systems. HPCA ’07, pp. 13–24, 2007.

[25] M. Ravishankar, J. Eisenlohr, L.-N. Pouchet, J. Ramanujam,A. Rountev, and P. Sadayappan. Code generation for parallelexecution of a class of irregular loops on distributed memorysystems. SC ’12, pp. 72:1–72:11, 2012.

[26] E. Schkufza, R. Sharma, and A. Aiken. Stochastic superopti-mization. ASPLOS ’13, pp. 305–316, 2013.

[27] A. M. Sloane. Lightweight language processing in Kiama.GTTSE III, pp. 408–425. Springer, 2011.

[28] M. M. Strout, L. Carter, and J. Ferrante. Compile-timecomposition of run-time data and iteration reorderings. PLDI’03, pp. 91–102, 2003.

[29] S. d. Swierstra and O. Chitil. Linear, bounded, functionalpretty-printing. J. Funct. Program., 19(1):1–16, Jan. 2009.

[30] V. K. Vavilapalli, A. C. Murthy, C. Douglas, S. Agarwal,M. Konar, R. Evans, T. Graves, J. Lowe, H. Shah, S. Seth,B. Saha, C. Curino, O. O’Malley, S. Radia, B. Reed, andE. Baldeschwieler. Apache Hadoop YARN: Yet anotherresource negotiator. SOCC ’13, pp. 5:1–5:16, 2013.

[31] M. Zaharia, M. Chowdhury, M. J. Franklin, S. Shenker, andI. Stoica. Spark: Cluster computing with working sets. Hot-Cloud’10, pp. 10–10, 2010.

A. WordCountfor ( i n t i = 0 ; i < docs . l e n g t h ; i ++) {

S t r i n g [ ] s p l i t = docs [ i ] . s p l i t ( " " ) ;for ( i n t j = 0 ; j < s p l i t . l e n g t h ; j ++) {

S t r i n g word = s p l i t [ j ] ;I n t e g e r p r ev = m. g e t ( word ) ;if ( p r ev == n u l l ) p r ev = 0 ;m. p u t ( word , p r ev + 1 ) ;

}}

+ Java to LambdaRange ( 0 , docs . s i z e ) . f o l d (m) ( {

c a s e ( v34 , i ) =>val s p l i t = docs ( i ) . s p l i t ( " " ) ;val v10 = s p l i t . s i z e ;Range ( 0 , v10 ) . f o l d ( v34 ) ( {

c a s e ( v33 , j ) =>val v13 = v33 ( s p l i t ( j ) ) ;val p rev = { if ( v13 != n u l l ) v13 else 0 } ;v33 . u p d a t e d ( s p l i t ( j ) , p r e v + 1)

} )} )

+ S:eliminate-null-checkRange ( 0 , docs . s i z e ) . f o l d (m) ( {

c a s e ( v34 , i ) =>val v6 = docs ( i ) . s p l i t ( " " ) ;val v4 = Range ( 0 , v6 . s i z e ) . groupBy ( {

c a s e ( j ) => v6 ( j )} ) ;

val v5 = v4 . map ( {c a s e ( v2 , v3 ) =>

v3 . f o l d ( v34 ( v2 ) ) ( {c a s e ( v1 , j ) => v1 + 1

} )} ) ;

v34 ++ v5} )

+ S:reduce-monoid-identity-opRange ( 0 , docs . s i z e ) . f o l d (m) ( {


c a s e ( j ) => v6 ( j )} ) ;

val v5 = v4 . map ( {c a s e ( v2 , v3 ) =>

v34 ( v2 ) + v3 . s i z e} ) ;

v34 ++ v5} )

+ E:localize-group-by-accessesRange ( 0 , docs . s i z e ) . f o l d (m) ( {

c a s e ( v34 , i ) =>val v4 = docs ( i ) . s p l i t ( " " ) . groupBy ( ID ) . __2 ;val v5 = v4 . map ( {

c a s e ( v2 , v3 ) =>v34 ( v2 ) + v3 . s i z e

} ) ;v34 ++ v5

} )

+ E:localize-map-accessesRange ( 0 , docs . s i z e ) . f o l d (m) ( {

c a s e ( v34 , i ) =>val v4 = docs ( i ) . s p l i t ( " " ) . groupBy ( ID ) . __2 ;val v5 = v4 . z i p ( v34 ) . map ( {

c a s e ( v2 , ( v3 , v7 ) ) =>v7 + v3 . s i z e

} ) ;v34 ++ v5

} )

[13] M. H. Hall, S. P. Amarasinghe, B. R. Murphy, S.-W. Liao,and M. S. Lam. Detecting coarse-grain parallelism using aninterprocedural parallelizing compiler. Supercomputing ’95,1995.

[14] R. Joshi, G. Nelson, and K. Randall. Denali: A goal-directedsuperoptimizer. PLDI ’02, pp. 304–314, 2002.

[15] R. A. Kelsey. A correspondence between continuation passingstyle and static single assignment form. IR ’95, pp. 13–22,1995.

[16] Y. Klonatos, A. Nötzli, A. Spielmann, C. Koch, and V. Kuncak.Automatic synthesis of out-of-core algorithms. SIGMOD ’13,pp. 133–144, 2013.

[17] K. Knobe and V. Sarkar. Array SSA form and its use inparallelization. POPL ’98, pp. 107–120, 1998.

[18] R. Lämmel. Google’s MapReduce programming model —revisited. Science of Computer Programming, 70(1):1 – 30,2008.

[19] S.-w. Liao. Parallelizing user-defined and implicit reductionsglobally on multiprocessors. ACSAC’06, pp. 189–202, 2006.

[20] E. Meijer, M. Fokkinga, and R. Paterson. Functional program-ming with bananas, lenses, envelopes and barbed wire. FPCA’91, pp. 124–144, 1991.

[21] C. Nugteren and H. Corporaal. Introducing Bones: a paralleliz-ing source-to-source compiler based on algorithmic skeletons.GPGPU-5, pp. 1–10, 2012.

[22] B. C. Oliveira, A. Moors, and M. Odersky. Type classes asobjects and implicits. OOPSLA ’10, pp. 341–360, 2010.

[23] N. Ramsey. Unparsing expressions with prefix and postfixoperators. Software: Practice and Experience, 28(12):1327–1356, 1998.

[24] C. Ranger, R. Raghuraman, A. Penmetsa, G. Bradski, andC. Kozyrakis. Evaluating MapReduce for multi-core andmultiprocessor systems. HPCA ’07, pp. 13–24, 2007.

[25] M. Ravishankar, J. Eisenlohr, L.-N. Pouchet, J. Ramanujam,A. Rountev, and P. Sadayappan. Code generation for parallelexecution of a class of irregular loops on distributed memorysystems. SC ’12, pp. 72:1–72:11, 2012.

[26] E. Schkufza, R. Sharma, and A. Aiken. Stochastic superopti-mization. ASPLOS ’13, pp. 305–316, 2013.

[27] A. M. Sloane. Lightweight language processing in Kiama.GTTSE III, pp. 408–425. Springer, 2011.

[28] M. M. Strout, L. Carter, and J. Ferrante. Compile-timecomposition of run-time data and iteration reorderings. PLDI’03, pp. 91–102, 2003.

[29] S. d. Swierstra and O. Chitil. Linear, bounded, functionalpretty-printing. J. Funct. Program., 19(1):1–16, Jan. 2009.

[30] V. K. Vavilapalli, A. C. Murthy, C. Douglas, S. Agarwal,M. Konar, R. Evans, T. Graves, J. Lowe, H. Shah, S. Seth,B. Saha, C. Curino, O. O’Malley, S. Radia, B. Reed, andE. Baldeschwieler. Apache Hadoop YARN: Yet anotherresource negotiator. SOCC ’13, pp. 5:1–5:16, 2013.

[31] M. Zaharia, M. Chowdhury, M. J. Franklin, S. Shenker, andI. Stoica. Spark: Cluster computing with working sets. Hot-Cloud’10, pp. 10–10, 2010.

A. WordCountfor ( i n t i = 0 ; i < docs . l e n g t h ; i ++) {

S t r i n g [ ] s p l i t = docs [ i ] . s p l i t ( " " ) ;for ( i n t j = 0 ; j < s p l i t . l e n g t h ; j ++) {

S t r i n g word = s p l i t [ j ] ;I n t e g e r p r ev = m. g e t ( word ) ;if ( p r ev == n u l l ) p r ev = 0 ;m. p u t ( word , p r ev + 1 ) ;

}}

+ Java to LambdaRange ( 0 , docs . s i z e ) . f o l d (m) ( {

c a s e ( v34 , i ) =>val s p l i t = docs ( i ) . s p l i t ( " " ) ;val v10 = s p l i t . s i z e ;Range ( 0 , v10 ) . f o l d ( v34 ) ( {

c a s e ( v33 , j ) =>val v13 = v33 ( s p l i t ( j ) ) ;val p rev = { if ( v13 != n u l l ) v13 else 0 } ;v33 . u p d a t e d ( s p l i t ( j ) , p r ev + 1)

} )} )

+ S:eliminate-null-checkRange ( 0 , docs . s i z e ) . f o l d (m) ( {


c a s e ( j ) => v6 ( j )} ) ;

val v5 = v4 . map ( {c a s e ( v2 , v3 ) =>

v3 . f o l d ( v34 ( v2 ) ) ( {c a s e ( v1 , j ) => v1 + 1

} )} ) ;

v34 ++ v5} )

+ S:reduce-monoid-identity-opRange ( 0 , docs . s i z e ) . f o l d (m) ( {


c a s e ( j ) => v6 ( j )} ) ;

val v5 = v4 . map ( {c a s e ( v2 , v3 ) =>

v34 ( v2 ) + v3 . s i z e} ) ;

v34 ++ v5} )

+ E:localize-group-by-accessesRange ( 0 , docs . s i z e ) . f o l d (m) ( {

c a s e ( v34 , i ) =>val v4 = docs ( i ) . s p l i t ( " " ) . groupBy ( ID ) . __2 ;val v5 = v4 . map ( {

c a s e ( v2 , v3 ) =>v34 ( v2 ) + v3 . s i z e

} ) ;v34 ++ v5

} )

+ E:localize-map-accessesRange ( 0 , docs . s i z e ) . f o l d (m) ( {

c a s e ( v34 , i ) =>val v4 = docs ( i ) . s p l i t ( " " ) . groupBy ( ID ) . __2 ;val v5 = v4 . z i p ( v34 ) . map ( {

c a s e ( v2 , ( v3 , v7 ) ) =>v7 + v3 . s i z e

} ) ;v34 ++ v5

} )

BrittleDifficult to be correct

Hard to maintain

SEARCH

Proof of translationTarget code

SEARCH

Proof of translationTarget code

DSL

<

Proof of translation

PROGRAM SYNTHESISTarget code

DSL

<

Program Compilation vs. Synthesis

Suppose we want to optimize x+x+x+x

With compiler rewrite rules:

With synthesis search:

x+x+x+x 4*x 2^2*x x<<2

x<<2

x*5x+3

x+2x

x<<42Space of programs

x-5

UnitTester

TheoremProver

ModelChecker

SpecificationLanguage

Infer Lift

Verified Lifting

Validation

CodeGeneration

Source Languages Target DSLs

Search

MetaLift

CodeIdentifier

SpecificationSynthesizer

TheoremProver

CodeGenerator

• Boolean• Arithmetic

• Classes• Lists• Arrays• …

DSL semanticsSearch space description

Target code fragments

Codegen rules

UnitTests

Compiler GeneratorSpec Language

MetaLift

CodeIdentifier


TheoremProver

CodeGenerator




Codegen rules


select :: [(a)] -> ((a)-> Bool) -> [(a)]select l f = [t | t <- l, f(t)]

join :: [(a)] -> [(b)] -> [(a,b)]join l1 l2 = [(t1, t2) | t <- l1, t <- l2]...

UnitTests

MetaLift

CodeIdentifier


TheoremProver

CodeGenerator



Codegen rules




Java: while (*) { * }

UnitTests

MetaLift

CodeIdentifier


TheoremProver

CodeGenerator







translate (select l f) = "select * from" + translate(l) + "where" + translate(f)...

UnitTests

Data Analytics: Java àSQL

List getUsersWithRoles () {List users = this.userDao.getUsers(); List roles = this.roleDao.getRoles(); List results = new ArrayList();for (User u : users) {

for (Role r : roles) {if (u.roleId == r.id)

results.add(u); }}

return results; }

List getUsersWithRoles () { return executeQuery(“SELECT u FROM users u, roles r WHERE u.roleId == r.idORDER BY u.roleId, r.id”);

}

Lifted code can be optimized by DBs100-1000x speedup

codegen

[PLDI 13, CIDR 14]

1.Defineordered listsandtheiroperations

2.Synthesizer infersspecfrom source

3.Retargetsynthesized spectoSQLselect :: [(a)] -> ((a)-> Bool) -> [(a)]select l f = [t | t <- l, f(t)]

join :: [(a)] -> [(b)] -> [(a,b)]join l1 l2 = [(t1, t2) | t <- l1, t <- l2]

Leveraging GPUs: Fortran à Halide

procedure sten(imin,imax,jmin,jmax,a,b) real,dim(imin:imax,jmin:jmax) :: a real,dim(imin:imax,jmin:jmax) :: b do j=jmin,jmax

t = b(imin, j) do i=imin+1,imax

q = b(i,j) a(i,j) = q + t t=q

enddoenddo

end procedure

int main() { ImageParam b(type_of<dbl>(),2); Func func; Var i, j; func(i,j) = b(i-1,j) + b(i,j); func.compile_to_file("ex1", b); return 0;

}

Lifted code can be executed on GPUs17x speedup

codegen

1.Definearraysand theiroperations


3.Retargetsynthesized spectoHalide

[SNAPL 15, PLDI 16]

get a i = a istore a i e = a//[(i,e)]

Hardware: Domino à Programmable Switches

void flowlet (Packet p) {p.new_hop = hash3(p.sport, p.dport,

p.arrival) % HOPS;p.id = hash2(p.sport, p.dport) % FLOWS;

if (p.arrival – last_time[p.id]>THRESHOLD)

saved_hop[pkt.id] = p.new_hop;

...}

Compile 10 well-known data plane algorithms to switches & run at line rate

codegen

1.Definehardwarebuilding blocksasinstructions


3.Retargetspectoprogrammable switches [SIGCOMM 16]mux c v1 v2 = if c then v1 else v2pairExec x y f g = (f x, g y)inc x e = x + e

Parallel Frameworks: Sequential Java à Hadoop

// sequential implementationint regress(Point [] points) {

int SumXY = 0;for (Point p : points){

SumXY += p.x * p.y;}

return SumXY;}

void map(Object key, Point [] value)

{ for (Point p : points)emit("sumxy", p.x * p.y);

}}

void reduce(Text key, int [] vs) { int SumXY = 0;

for (Integer val : vs)SumXY = SumXY + val;

emit(key, SumXY); }

Lifted code can be optimized by Hadoop/Spark32x speedup

codegen

1.Definesemanticsofmapandreduce


3.Retargetsynthesized spectoHadoop andSparkmap l f = [f t | t <- l] reduce l f i = foldl f i l

[SYNT 16, SIGMOD 17]

CasperCompiling Sequential Java to HadoopAn Illustration of the MetaLift Toolchain

CodeIdentifier


TheoremProver

CodeGenerator

Casper Architecture



Codegen rules

CodeIdentifier


TheoremProver

CodeGenerator

Casper Architecture


Codegen rules


map :: [(a)] -> (a -> (k,v)) -> [(k,v)]map l f = [f t | t <- l]

reduce :: [(k,[v])] -> (v -> v -> v) -> [(k,v)]reduce l f i = [(k, foldl f i v) | (k,v) <- l]

translate (map l f) ="void map(...) { for (t : l) emit(" + translate(f) + "; }"

translate (reduce l f i) = ...

DSL semantics

Code generation

Search Space Grammar

e := (e1, e2) | e + e | e - e | var | lit

| e * e | e / e | if e1 then e2 else e3

Automatically generated for each input code fragment

Increases expressivity incrementally

CandidateSpecGenerator

BoundedModelChecker

Search space grammar

TheoremProver

Candidate Spec

SynthesisCoordinatorInput code fragment

Correct Specifications

Candidate Spec

CodeIdentifier


TheoremProver

CodeGenerator

Evaluation: Benchmarks

60 benchmarks collected from 5 benchmark suites

Phoenix: Classic MapReduce problems such as WordCount, 3D Histogram, Linear Regression and StringMatch.

Bigλ: Big-Data analytical benchmarks including basic sentiment analysis, database operations and Wikipedia log processing.

Arithmetic: Mathematical functions such as Max, Delta and Conditional Sum.

Statistical: Statistical analysis such as Mean, Covariance and Standard Error.

Fiji: Three image processing kernels: RedToMagenta, Temporal Median and Trails.

Feasibility: Does Casper work?Benchmark Extracted Translated

Phoenix 7 11

Bigλ 6 8

Arithmetic 11 11

Statistical 18 19

Fiji 8 11

Total 50 60

Causes of failures• 3 caused by calls to unsupported to external library methods• 7 caused by program constructs not currently expressible using MetaLift

Performance: Is Casper competitive?

0x

5x

10x

15x

20x

25x

30x

35x

StringMatch WordCount LinearRegression

3DHistogram YelpKids WikipediaPageCount

Covariance HadamardProduct

DatabaseSelect

AnscombeTransform

Speedup

MOLD

Spark(Casper)

Mean speedup: 17.6x Max speedup: 31x

0x

5x

10x

15x

20x

25x

30x

35x

StringMatch WordCount LinearRegression

3DHistogram YelpKids WikipediaPageCount

Covariance HadamardProduct

DatabaseSelect

AnscombeTransform

Speedup

MOLDSpark(Manual)Spark(Casper)

Competitive with expert implementations

Performance: Is Casper competitive?

Performance Analysis: Is Casper Practical?

Mean compilation time for one benchmark was 5.8 minutes.

Median compilation time for one benchmark was 36.6 seconds.

Mean time to reach first correct translation was only 48 seconds!

BenchmarkSolutions Generated

with PruningSolutions Generated

without Pruning

WordCount 2 827

3D Histogram 5 118

Yelp: Kid Friendly Restaurants 1 286

LEVERAGING DSLS (ALMOST) FOR FREE @alvinkcheungmetalift.uwplse.org/strangeloop17.pdf · [17] K....

Documents

Transcript of LEVERAGING DSLS (ALMOST) FOR FREE @alvinkcheungmetalift.uwplse.org/strangeloop17.pdf · [17] K....