R FUTURE DIRECTIONS. Speaker. FastR (speedup 8.5x GNURast) What have we done so far. BRIXEN June Jan Vitek. Professor of Computer Science

Size: px

Start display at page:

Download "R FUTURE DIRECTIONS. Speaker. FastR (speedup 8.5x GNURast) What have we done so far. BRIXEN June Jan Vitek. Professor of Computer Science"

Hugo Page
5 years ago
Views:

R FUTURE DIRECTIONS Jan Vitek Speaker Professor of Computer Science Purdue University, USA PHD: University of Geneva, CH BRIXEN June 2014 Speciality: Programming languages, Compilers What have we

1 R FUTURE DIRECTIONS Jan Vitek Speaker Professor of Computer Science Purdue University, USA PHD: University of Geneva, CH BRIXEN June 2014 Speciality: Programming languages, Compilers What have we done so far FastR (speedup 8.5x GNURast) TracR a trace analysis / profiler for GNU-R CoreR a formal semantics for a fragment of R TestR a testing framework for the R language FastR a new R virtual machine written in Java Morandat, Hill, Osvald, Vitek. Evaluating the Design of the R Language. ECOOP 12 Relative Execution Time (GNUR AST = 1) FastR GNUR BC Kalibera, Maj, Morandat, Vitek. A Fast abstract syntax tree interpreter for R. VEE 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 Figure 13. Shootout Relative Execution Times

2 Evaluating the Design of R R S-1 S-2 S-3 S-4 S-5 S-6 S-7 S-8 S-9 S- S-11 S Python 15 Name Binary trees Fankuch redux Fasta Fasta redux K-nucleotide Mandelbrot N-body Pidigits Regex-dna Rev. complement Spectral norm Spectral norm alt Input 16 50K 4K K 2.5K 5M K Avg S 12 S 11 S 9 S S 8 S 7 S 6 S 5 S 4 S 3 S 2 S PERFORMANCE Fig. 6. Slowdown of Python and R, normalized to C for the Shootout benchmarks. Evaluating the Design of R Python R To understand where the Name Input time is spent, we turn to more S-1 Fig. Binary 16 representative R programs. 7 trees shows the breakdowns-2 of execufankuch redux tion times in the Bioconductor S-3 Fasta mm dataset obtained withs-4 ProfileR. Fasta redux alloc.cons Each bar represents S-5 a BioconK-nucleotide 50K alloc.list alloc.vector ductor vignette. The S-6 key obsermandelbrot 4K duplicate vation is that memory managelookup S-7 N-body K match ment accounts for an S-8 average of Pidigits external 29% of execution time. Memory builtin S-9 Regex-dna 2.5K arith management time was special S-further Rev. complement 5M broken down into time spent in S-11 Spectral norm 640 garbage collection (18.7%), als-12 Spectral norm alt 11K locating cons-pairs (3.6%), vectors (2.6%), and duplications Performance relative to C (4%) for call-by-value semanslowdown of Python and R, normalized to C for the Shootout benchmarks. tics. The time spent in builtshootout Time breakdown in functions represents the true Fig. 7. Time breakdown of Bioconductor vignettes. computational work performed by R, this is on average 38% of execution time. There are some interesting outliers Avg S 12 S 11 S S 9 S 8 S 7 S 6 S 5 S 4 S 3 S 2 S Fig

3 SPECIFICATION C R User data R internal S 1 S 2 S 3 S 4 S 5 S 6 S 7 S 8 Heap Memory Shootout S 9 S S 11 S 12 45% of assignments are definitions cpy(h, ) =H 0, 0 = 0 H( ) =F [A F 0 = F [x/ 0 ] H 00 = H 0 [ /F 0 ] 0 0 x < ; H ; H 00 only 2 out of 217 million assignments are definitions in a parent frame 99.9% of side effects are local

4 c <- 42 c <- 42 d <- c c(1,2,3) d(1,2,3) [1] Error: could not find function "d" less than 0.05% context sensitive function name lookups only symbols that rely on it are c and file OBJECTS Bioc Misc CRAN Base Total # classes # methods S3 Avg. redef Method calls 13M 58M M Super calls 697K 1.2M - - 2M # classes # singleton # leaves Hier. depth S4 Direct supers # methods Avg. redef Redef. depth # new 668K K Method calls 15M M Super calls 94K K

5 x <- y x : like D[20,20] <- y TYPES x : D[20,20] <- y x : D[ dim(y) ] <- y 1:00000 DATA LAYOUT x[[ ]]

6 x <<- 1 x <<- 3 CONCURRENCY Effort Implem. Project language start/end KLOC Man/ month Percent Complete Missing parts Challenges Time sinks GNUR C (Fortran) +? 99.9??? RIPOSTE C Lang Libs Std libs, pkgs, search paths, locked envs, FFIs Too large internal surface, src/main too big, messy headers, Fortran RENJIN Java 20 -? 8 78 primitives S4 exports, R compatibility, native code JIT, data layout, C->R internals IMPLEMENTATIONS TERR C ORBIT C (Fortran) R compatibility, Graphics Some bugs Clean-room imp Native code Complexity of runtime S3,S4, data.frame, model.frame, native interface Manual specializations FastR.168 Java / graphics, internals, R compatibility, native interface, Internals, manual specialization, Truffle, benchmarks FastR Java Lang 49 Libs R compatibility, Graphics, promises, frames/envs R compatibility, vector ops, frames/envs.

7 the road ahead

A Fast Abstract Syntax Tree Interpreter for R

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