6 Speeding up MARSS and allowing larger matrices
date: May 3, 2018
Problem:
- MARSS is slow in part due to native R
- MARSS is very slow with large matrices
I am working on an update to MARSS to allow sparse matrices to allow larger datasets to be used and Rcpp integration to speed up large matrix math.
Install current version of MARSS into one location.
library(devtools)
withr::with_libpaths(.libPaths()[2], install_github("nwfsc-timeseries/MARSS@*release"))
Install Matrix branch of MARSS into one location.
library(devtools)
withr::with_libpaths(.libPaths()[1], install_github("nwfsc-timeseries/MARSS", ref="Matrix"))
Functions to fit models.
fit.vs9=function(n,TT){
library(MARSS, lib.loc=.libPaths()[2])
set.seed(111)
a = matrix(rnorm(n*TT),n,TT)
kemfit=MARSS(a, control=list(maxit=20), fun.kf="MARSSkfss", silent=TRUE, fit=FALSE)
tmp=MARSSkem(kemfit)
detach("package:MARSS", unload=TRUE)
return(tmp)
}
fit.Matrix=function(n,TT){
library(MARSS, lib.loc=.libPaths()[1])
set.seed(111)
a = matrix(rnorm(n*TT),n,TT)
isMatrix=TRUE # for now I use a flag to denote to use sparse Matrix
kemfit=MARSS(a, control=list(maxit=20), fun.kf="MARSSkfss", silent=TRUE, fit=FALSE)
tmp=MARSSkem(kemfit)
detach("package:MARSS", unload=TRUE)
return(tmp)
}
Now benchmark. This takes awhile
library(microbenchmark)
microbenchmark(fit.vr9(75,20), fit.Matrix(75,20))
microbenchmark(fit.vr9(100,20), fit.Matrix(100,20))
microbenchmark(fit.vr9(200,20), fit.Matrix(200,20))
Looks similar but the difference is in the MARSSkem() and memory. The set-up has not been optimized for Matrix form yet and takes much longer. You can see this best with the Rstudio profiler.
library(profvis)
profvis({ tmp=fit.Matrix(100,20) })
profvis({ tmp=fit.vs9(100,20) })
Time is 17.5 sec (Matrix) versus 16.8 sec (vs 9). But memory is 6G vs 10.8G. The MARSSkem part is 8.8 sec vs 12 sec, but that is from the C++ speeds ups and will be transferred to the non-Matrix version too. Object sizes are much smaller for the Matrix version, 18M vs 0.33M.
With 200x20 matrix, the differences are greater. 97 sec vs 182 sec.