optimize sim.outer.tree()

[ArtPoon]

This is taking up most of the time:

types <- sapply(active, function(comp) comp$get.type()$get.name())

[ArtPoon]

[ArtPoon]

I tried generating types once as a list (lapply) and then updating it as we update active, but this caused .assign.events to fail.

[ArtPoon]

--- a/R/simOuterTree.R
+++ b/R/simOuterTree.R
@@ -229,8 +229,8 @@ sim.outer.tree <- function(model, max.attempts=5) {
       break
     }
     
-    attempt <- attempt+1  # else try again
-    warning("Failed to resolve outer tree, starting attempt ", attempt)
+    attempt <- attempt + 1  # else try again
+    warning("Failed to resolve outer tree, starting attempt ", attempt, immediate.=TRUE)
   }
   
   if (attempt == max.attempts) {
@@ -548,6 +548,7 @@ sim.outer.tree <- function(model, max.attempts=5) {
   
   # start with sampled infected Compartments
   active <- run$get.compartments()
+  types <- lapply(active, function(comp) comp$get.type()$get.name())
   
   # iterate through events in reverse time (start with most recent)

[ArtPoon]

Fixed the above problem, code now runs faster but now rbind is consuming time:

Pre-allocate containers to avoid calling rbind repeatedly.

[ArtPoon]

Lots faster now!

[ArtPoon]

Consider using data.table for EventLogger:
https://rstudio-pubs-static.s3.amazonaws.com/406521_7fc7b6c1dc374e9b8860e15a699d8bb0.html

[ArtPoon]

There's still a lot of time spent doing garbage collection <GC> in .sample.outer.events - check what is being dynamically allocated.

[ArtPoon]

This is the current testing code:

require(twt)
require(profvis)
require(microbenchmark)

path <- system.file('extdata', 'structSI.yaml', package='twt')
settings <- yaml.load_file(path)

settings$InitialConditions$originTime <- 15
settings$InitialConditions$size$host1 <- 5000
settings$InitialConditions$size$host2 <- 5000

settings$CompartmentTypes$host1$effective.size <- 100
settings$CompartmentTypes$host1$branching.rates <- '(host1=1e-4, host2=2e-5)'
settings$CompartmentTypes$host1$migration.rates <- '(host1=5e-5, host2=1e-5)'
settings$CompartmentTypes$host1$transition.rates <- '(host1=0, host2=1e-2)'

settings$CompartmentTypes$host2$effective.size <- 100
settings$CompartmentTypes$host2$branching.rates <- '(host1=2e-5, host2=1e-4)'
settings$CompartmentTypes$host2$migration.rates <- '(host1=1e-5, host2=5e-5)'
settings$CompartmentTypes$host2$transition.rates <- '(host1=1e-2, host2=0)'

settings$Compartments$I1$replicates <- 500
settings$Compartments$I2$replicates <- 500

model <- Model$new(settings)

#profvis({
#  run <- sim.outer.tree(model)
#})


timing <- sapply(1:100, function(i) {
  print(i)
  t0 <- proc.time()[3]
  run <- sim.outer.tree(model)
  proc.time()[3] - t0
})

[ArtPoon]

One possibility is to replace row assignment of data frame with matrix operations:

twt/R/simOuterTree.R

Lines 484 to 489 in 2942e1f

	# append counts to outcome container after event
	#counts <- rbind(counts, c(susceptible, infected))
	counts[row.num, ] <- c(susceptible, infected)
	# append event
	events[row.num, ] <- c(current.time, event.type, recipient$get.name(), source$get.name())

See https://stackoverflow.com/questions/48703802/most-efficient-way-fastest-to-modify-a-data-frame-using-indexing