A Fast Abstract Syntax Tree Interpreter for R

Transcription

1 A Fast Abstract Syntax Tree Interpreter for R Java cup Petr Maj Jan Vitek Tomas Kalibera Floréal Morandat Thesis Runtime information can be leveraged to create simple, fast, easy to maintain interpreters for real languages Purdue University I 0 7 % ata d of R r Tools Used feby Data Miners e pr s rtools Used by Data Miners e n mi Vendors were excluded from these analyses Vendors were excluded from these analyses Into the data mines By Rexer Analytics, presented at Predictive Analytics World Boston, 2013 By Rexer Analytics, presented at Predictive Analytics World Boston, respondents from 75from countries 1259 respondents 75 countries

2 1976 S 1993 R Today, The R project An R history Bell Labs, then S-Plus (closed-source owned by Tibco) Ihaka and Gentlemann, started R as new language at the U of Auckland, NZ Core team ~ 20 people, under GPL. Continued dev of language & libraries: namespaces ( 11), bytecode ( 11), 64-bit indexes ( 13) Vectors x <- c(2,7,9,na,5) c(1,2,3) + x[1:3] <-function(data,,...) x[is.na(x)] <- 0 Functions p<-function(x=5,,y=x+1) x + c( ) + y p() p(1,2) p(y=2) <-function(data,,...) p(y=2,x=1) p(1,2,y=0,3,4) Promises assert(x==2, report(x)) <-function(data,,...) assert<-function(c,p) if(c)print(p)

3 Referential transparency f(x) x <- c(0,1) assert(x[[1]]==1) <-function(data,,...) f<-function(a){a[[1]]<-0} Reflection with(fd, carb*den) with<-function(data,, ) eval(substitute(), <-function(data,,...) Attributes x <- c(1,2,3,4) attr(x,"dim")<-c(2,2) <-function(data,,...) II Of interpreters and men

4 Performance & Complexity Basic textual interpreter Abstract syntax tree interpreter Bytecode interpreter JIT compiler Tracing JIT compiler Optimizing JIT DLS 12 Self-Optimizing AST Interpreters Thomas Wurthinger, Andreas Woß, Lukas Stadler, Gilles Duboscq, Doug Simon, Christian Wimmer Showed how to obtain performance for a simple(r) language We use Truffle as an inspiration to develop techniques for R Renjin: R in Java 1.8x slower III 1 0 bintree fannkrdx fasta fastardx knucleo knucleo-br mandel mandel-nv nbody pidigits regexdna reverse reverse-nv spectral spectral-alt spectral-mth Worse is better

5 Preparation Code Specialization In-place profile-driven code specialization ANTLR Parse Tree Optimizations: allocate slot for a subset of the local variables return statement elision Only compilation step in FastR FastR Executable Tree N 1 + program state N N 2 N 1 N 2 Use values of variables to choose a better implementation Simple to implement in single threaded context, modulo some care around recursive functions + program state + exception? N Code Specialization Code Specialization f(12, x+1, a=3) gen_call f f <- function(a,b,c){a+c} pos_call f <- function(a,b,c){a+c} 12 x+1 3 a 3 12 x+1

6 Code Specialization class If { RNode conde, trueb, falseb; Object execute(frame f) { try { val = conde.executescalarlogical(frame); } catch (UnectedResult e) { cast = ToLogical.mkNode(condE, e.result()); replacechild(conde, cast); return execute(frame); } if (val == TRUE) return trueb.execute(f); if (val == FALSE) return falseb.execute(f); throw unectedna(); } Code Specialization Specialized code for simple cases, e.g. arithmetics with scalars; function calls scalar vector indexing arithmetics on vectors of same length Bounded self-rewriting: - Uninitialized node rewrites itself - Initialized node rewrites itself when a guard fails Rewriting based on events: e.g. a symbol change can re-writes nodes Data Specialization Data Specialization A special representation for common cases of data-types Scalar integer with no names, no dimensions, no custom attributes... just the number Integer ranges A specialized implementation of an integer vector which happens to be a range 1,2,,k k N, k>0 Saves memory and enables code specialization Reduces memory overhead Merges guards needed in (specialized) code if (v instanceof RInt && v.size() == 1 && v.names() == null && v.dimensions() == null && v.attributes() == null) {... for(i in 1:n) {...} Java loop over Java integers 1 to n x[,1:n] = NA Column selection using a Java loop from 1 to n if (v instanceof ScalarIntImpl) {...

7 Data Specialization class RIntSimpleRange implements RInt { Integer ranges final int to;" } int getint(int i) { return i + 1; }" RInt materialize() { " int[] c = new int[to]; " for (int i = 0; i < to; i++) c[i] = i + 1; return RInt.RIntFactory.getFor(c);" }" A range trivially won't contain an NA, negative values, zero, and bounds check for whole range is constant time Much faster vector indexing Views Data+Code Specialization a <- b+c o <- a*2 * delay construction of large data objects are a data-flow representation of vectors avoid unnecessary work if a subset of the data is required avoid allocation of temporary objects permit fusion of multiple data traversals into one Data+Code Specialization FastR (speedup 8.5x) a <- b+c P FastR GNUR BC Profiling view 0.75 Counts number of accesses to a view An executable node, on first invocation, creates a profiling view; second invocation, rewrites based on the profile 0.50 Heuristics: 0.25 Avoid small views; avoid over-used root views; avoid repeating computation; avoid recursive views 0.00 bt1 bt2 bt3 pr1 pr2 fa1 fa2 fa3 fa4 fa5 fr1 fr2 kn1 kn2 kn3 kn4 ma1 ma2 ma3 ma4 nb1 nb2 nb3 nb4 nb5 pd1 rd1 rc1 rc2 rc3 sn1 sn2 sn3 sn4 sn5 sn6 sn7

8 Conclusion Runtime information can be leveraged to create simple, fast, easy to maintain interpreters for real languages