In CamFreshwater/samSim: Closed-loop salmon dynamics simulator

Conduct some trial simulation runs to try to improve run times.

listOfPackages <- c("plyr", "here", "parallel", "doParallel", "foreach", 
                    "reshape2", "tidyverse", "tictoc", "synchrony", "zoo", 
                    "Rcpp", "RcppArmadillo", "bench", "samSim", "profvis")

here <- here::here

newPackages <- listOfPackages[!(listOfPackages %in% 
                                  installed.packages()[ , "Package"])]
if(length(newPackages)) install.packages(newPackages)
lapply(listOfPackages, require, character.only = TRUE)

simPar <- read.csv(here("data/manProcScenarios/fraserMPInputs_varyAllocationVaryFixedER.csv"), 
                   stringsAsFactors = F)
# simPar <- read.csv(here("data/opModelScenarios/fraserOMInputs_varyCorr.csv"), 
#                    stringsAsFactors = F)
cuPar <- read.csv(here("data/fraserDat/fraserCUpars.csv"), stringsAsFactors = F)
srDat <- read.csv(here("data/fraserDat/fraserRecDatTrim.csv"), stringsAsFactors = F)
catchDat <- read.csv(here("data/fraserDat/fraserCatchDatTrim.csv"), 
                     stringsAsFactors = F)
ricPars <- read.csv(here("data/fraserDat/pooledRickerMCMCPars.csv"), stringsAsFactors = F)
larkPars <- read.csv(here("data/fraserDat/pooledLarkinMCMCPars.csv"), stringsAsFactors = F)
tamFRP <- read.csv(here("data/fraserDat/tamRefPts.csv"), stringsAsFactors = F)


nTrials <- 10
## General robustness runs
# simParTrim <- subset(simPar,
#                      scenario == "lowSig" | scenario == "medSig" |
#                        scenario == "highSig")
simParTrim <- simPar[81,]

scenNames <- unique(simParTrim$scenario)
dirNames <- sapply(scenNames, function(x) paste(x, unique(simParTrim$species),
                                                sep = "_"))

First run a profiler to try to identify likely bottlenecks.

profvis({
  recoverySim(simParTrim[1, ], cuPar, catchDat = catchDat, srDat = srDat, 
                variableCU = FALSE, ricPars, larkPars = larkPars, 
                tamFRP = tamFRP, cuCustomCorrMat = NULL, dirName = dirNames[1], 
                nTrials = 10, makeSubDirs=TRUE, random = FALSE)
})

Even with quickLM function there was a major bottleneck when calculating CU-specific trends using geometric means. These are useful BMs but not always used. Added two if statements if (!is.null(simPar$statusTrendPM)) and if (simPar$statusTrendPM == TRUE) so that they are only calculated when needed. Otherwise no obvious bottlenecks besides plotting diagnostics (which only happens once).

# Single core 
tic()
  for(i in seq_along(dirNames)) {
    d <- subset(simParTrim, scenario == scenNames[i])
    simsToRun <- split(d, seq(nrow(d)))
    lapply(simsToRun, function(x) {
      recoverySim(x, cuPar, catchDat=catchDat, srDat=srDat, variableCU=FALSE,
                  ricPars, larkPars=larkPars, tamFRP=tamFRP, dirName=dirNames[i],
                  nTrials=nTrials, makeSubDirs=TRUE, random = FALSE)
    })
  }
toc()

# Six cores
tic()
for (i in seq_along(dirNames)) {
  dirName <- dirNames[i]
  d <- subset(simParTrim, scenario == scenNames[i])
  simsToRun <- split(d, seq(nrow(d)))
  Ncores <- detectCores()
  cl <- makeCluster(Ncores - 2) #save two cores
  registerDoParallel(cl)
  clusterEvalQ(cl, c(library(MASS),
                  library(here),
                  library(sensitivity),
                  library(mvtnorm),
                  library(scales), #shaded colors for figs
                  library(viridis), #color blind gradient palette
                  library(gsl), 
                  library(dplyr),
                  library(Rcpp),
                  library(RcppArmadillo),
                  library(sn),
                  library(samSim)))
  #export custom function and objects
  clusterExport(cl, c("simsToRun", "recoverySim", "cuPar", "dirName", "nTrials",
                      "catchDat", "srDat", "ricPars", "dirName", "larkPars", 
                      "tamFRP"), envir = environment()) 
  parLapply(cl, simsToRun, function(x) {
    recoverySim(x, cuPar, catchDat = catchDat, srDat = srDat, 
                variableCU = FALSE, ricPars, larkPars = larkPars, 
                tamFRP = tamFRP, cuCustomCorrMat = NULL, dirName = dirName, 
                nTrials = nTrials, makeSubDirs=TRUE, random = FALSE)
  })
  stopCluster(cl) #end cluster
}
toc()

Running in parallel is ~150% faster even with low number of trials - relative gains will increase since there is some time associated with ramping up and shutting down cores.

CamFreshwater/samSim documentation built on Sept. 25, 2023, 10:22 a.m.