Skip to contents

Saving simulation results and state

Phil Chalmers

2024-11-14

Source: vignettes/Saving-results.Rmd
Saving-results.Rmd

Unsurprisingly, you may want to save your results to your hard disk in case of power outages or random system crashes to allow restarting at the interrupted location, save more complete versions of the analysis results in case you want to inspect the complete simulation results at a later time, store/restore the R seeds for debugging and replication purposes, and so on. This document demonstrates various ways in which SimDesign saves output to hard disks.

As usual, define the functions of interest.

library(SimDesign)
# SimFunctions()

Design <- createDesign(N = c(10,20,30))
Generate <- function(condition, fixed_objects) {
    dat <- rnorm(condition$N)    
    dat
}

Analyse <- function(condition, dat, fixed_objects) {
    ret <- c(p = t.test(dat)$p.value)
    ret
}

Summarise <- function(condition, results, fixed_objects) {
    ret <- EDR(results, alpha = .05)
    ret
}

This is a very simple simulation that takes very little time to complete, however it will be used to show the basic saving concepts supported in SimDesign. Note that more detailed information is located in the runSimulation documentation.

Option: save = TRUE (Default is TRUE)

The save flag triggers whether temporary results should be saved to the hard-disk in case of power outages and crashes. When this flag is used results can easily be restored automatically and the simulation can continue where it left off after the hardware problems have been dealt with. In fact, no modifications in the code required because runSimulation() will automatically detect temporary files to resume from (so long as they are resumed from the same computer node; otherwise, see the save_details list).

As a simple example, say that in the N=30N=30 condition something went terribly wrong and the simulation crashed. However, the first two design conditions are perfectly fine. The save flag is very helpful here because the state is not lost and the results are still useful. Finally, supplying a filename argument will safely save the aggregate simulation results to the hard-drive for future reference; however, this won’t be called until the simulation is complete.

Analyse <- function(condition, dat, fixed_objects) {
    if(condition$N == 30) stop('Danger Will Robinson!')
    ret <- c(p = t.test(dat)$p.value)
    ret
}

res <- runSimulation(Design, replications = 1000, save=TRUE, filename='my-simple-sim',
                     generate=Generate, analyse=Analyse, summarise=Summarise,
                     control = list(stop_on_fatal = TRUE))
## 
## Design: 1/3;   Replications: 1000;   RAM Used: 67.8 Mb;   Total Time: 0.00s 
##  Conditions: N=10
## 
## Design: 2/3;   Replications: 1000;   RAM Used: 68.8 Mb;   Total Time: 0.18s 
##  Conditions: N=20
## 
## Design: 3/3;   Replications: 1000;   RAM Used: 68.8 Mb;   Total Time: 0.34s 
##  Conditions: N=30
##

Check that temporary file still exists.

files <- dir()
files[grepl('SIMDESIGN', files)]
## [1] "SIMDESIGN-TEMPFILE_fv-az520-999.rds"

Notice here that the simulation stopped at 67% because the third design condition threw too many consecutive errors (this is a built-in fail-safe in SimDesign). To imitate a type of crash/power outage, control = list(stop_on_fatal = TRUE) input; otherwise, the simulation would continue normally over these terminal conditions though place NA placeholders for the terminal condition.

After we fix this portion of the code the simulation can be restarted at the previous state and continue on as normal. Therefore, in the event of unforeseen program execution crashes no time is lost.

Analyse <- function(condition, dat, fixed_objects) {
    ret <- c(p = t.test(dat)$p.value)
    ret
}

res <- runSimulation(Design, replications = 1000, save=TRUE, filename='my-simple-sim',
                     generate=Generate, analyse=Analyse, summarise=Summarise)
## 
## Design: 3/3;   Replications: 1000   Total Time: 0.34s 
##  Conditions: N=30
##

Check which files exist.

files <- dir()
files[grepl('SIMDESIGN', files)]
## character(0)
files[grepl('my-simp', files)]
## [1] "my-simple-sim.rds"

Notice that when complete, the temporary file is removed from the hard-drive.

Relatedly, the .Random.seed states for each successful replication can be saved by passing control = list(store_Random.seeds = TRUE)) to , though these are generally only useful under exceptional circumstances (e.g., when the generate-analyse results are unusual but did not throw a warning or error message, yet should be inspected interactively).

Passing store_results = TRUE stores the results object information that are passed to Summarise() in the returned object. This allows for further inspection of the simulation results, and potential to use functions such as reSummarise() to provide meta-summaries of the simulation at a later time. After the simulation is complete, these results can be extracted using SimResults(res) (or more generally with SimExtract(res, what = 'results')). For example,

# store_results=TRUE by default
res <- runSimulation(Design, replications = 3, 
              generate=Generate, analyse=Analyse, summarise=Summarise)
## 
## Design: 1/3;   Replications: 3;   RAM Used: 69.5 Mb;   Total Time: 0.00s 
##  Conditions: N=10
## 
## Design: 2/3;   Replications: 3;   RAM Used: 69.4 Mb;   Total Time: 0.00s 
##  Conditions: N=20
## 
## Design: 3/3;   Replications: 3;   RAM Used: 69.5 Mb;   Total Time: 0.01s 
##  Conditions: N=30
## 
results <- SimResults(res)
results
## # A tibble: 9 × 2
##       N     p
##   <dbl> <dbl>
## 1    10 0.363
## 2    10 0.816
## 3    10 0.555
## 4    20 0.674
## 5    20 0.481
## 6    20 0.120
## 7    30 0.664
## 8    30 0.527
## 9    30 0.848

Note that this should be used if the number of replications/design conditions is small enough to warrant such storage; otherwise, the R session may run out of memory (RAM) as the simulation progresses. Otherwise, save_results = TRUE described below is the recommended approach to resolve potential memory issues.

Finally, the save_results argument will output the results elements that were passed to Summarise() to separate .rds files containing all the analysis results and condition information. This option is supported primarily for simulations that are anticipated to have memory storage issues, where the results are written to the hard-drive and released from memory. Note that when using save_results the save flag is automatically set to TRUE to ensure that the simulation state is correctly tracked.

res <- runSimulation(Design, replications = 1000, save_results=TRUE,
                     generate=Generate, analyse=Analyse, summarise=Summarise)
## 
## Design: 1/3;   Replications: 1000;   RAM Used: 71.7 Mb;   Total Time: 0.00s 
##  Conditions: N=10
## 
## Design: 2/3;   Replications: 1000;   RAM Used: 71.7 Mb;   Total Time: 0.17s 
##  Conditions: N=20
## 
## Design: 3/3;   Replications: 1000;   RAM Used: 71.7 Mb;   Total Time: 0.33s 
##  Conditions: N=30
## 
dir <- dir()
directory <- dir[grepl('SimDesign-results', dir)]
dir(directory)
## [1] "results-row-1.rds" "results-row-2.rds" "results-row-3.rds"

Here we can see that three .rds files have been saved to the folder with the computer node name and a prefixed 'SimDesign-results' character string. Each .rds file contains the respective simulation results (including errors and warnings), which can be read in directly with readRDS():

row1 <- readRDS(paste0(directory, '/results-row-1.rds'))
str(row1)
## List of 6
##  $ condition    : tibble [1 × 1] (S3: tbl_df/tbl/data.frame)
##   ..$ N: num 10
##  $ results      :'data.frame':   1000 obs. of  1 variable:
##   ..$ p: num [1:1000] 0.976 0.397 0.143 0.595 0.453 ...
##  $ errors       : 'table' int[0 (1d)] 
##   ..- attr(*, "dimnames")=List of 1
##   .. ..$ : NULL
##  $ error_seeds  : NULL
##  $ warnings     : 'table' int[0 (1d)] 
##   ..- attr(*, "dimnames")=List of 1
##   .. ..$ warnings: NULL
##  $ warning_seeds: NULL
row1$condition
## # A tibble: 1 × 1
##       N
##   <dbl>
## 1    10
head(row1$results)
##        p
## 1 0.9759
## 2 0.3974
## 3 0.1430
## 4 0.5947
## 5 0.4534
## 6 0.2007

or, equivalently, with the SimResults() function

# first row
row1 <- SimResults(res, which = 1)
str(row1)
## List of 6
##  $ condition    : tibble [1 × 1] (S3: tbl_df/tbl/data.frame)
##   ..$ N: num 10
##  $ results      :'data.frame':   1000 obs. of  1 variable:
##   ..$ p: num [1:1000] 0.976 0.397 0.143 0.595 0.453 ...
##  $ errors       : 'table' int[0 (1d)] 
##   ..- attr(*, "dimnames")=List of 1
##   .. ..$ : NULL
##  $ error_seeds  : NULL
##  $ warnings     : 'table' int[0 (1d)] 
##   ..- attr(*, "dimnames")=List of 1
##   .. ..$ warnings: NULL
##  $ warning_seeds: NULL

The SimResults() function has the added benefit that it can read-in all simulation results at once (only recommended if RAM can hold all the information), or simply hand pick which ones should be inspected. For example, here is how all the saved results can be inspected:

input <- SimResults(res)
str(input)
## List of 3
##  $ :List of 6
##   ..$ condition    : tibble [1 × 1] (S3: tbl_df/tbl/data.frame)
##   .. ..$ N: num 10
##   ..$ results      :'data.frame':    1000 obs. of  1 variable:
##   .. ..$ p: num [1:1000] 0.976 0.397 0.143 0.595 0.453 ...
##   ..$ errors       : 'table' int[0 (1d)] 
##   .. ..- attr(*, "dimnames")=List of 1
##   .. .. ..$ : NULL
##   ..$ error_seeds  : NULL
##   ..$ warnings     : 'table' int[0 (1d)] 
##   .. ..- attr(*, "dimnames")=List of 1
##   .. .. ..$ warnings: NULL
##   ..$ warning_seeds: NULL
##  $ :List of 6
##   ..$ condition    : tibble [1 × 1] (S3: tbl_df/tbl/data.frame)
##   .. ..$ N: num 20
##   ..$ results      :'data.frame':    1000 obs. of  1 variable:
##   .. ..$ p: num [1:1000] 0.824 0.823 0.857 0.904 0.297 ...
##   ..$ errors       : 'table' int[0 (1d)] 
##   .. ..- attr(*, "dimnames")=List of 1
##   .. .. ..$ : NULL
##   ..$ error_seeds  : NULL
##   ..$ warnings     : 'table' int[0 (1d)] 
##   .. ..- attr(*, "dimnames")=List of 1
##   .. .. ..$ warnings: NULL
##   ..$ warning_seeds: NULL
##  $ :List of 6
##   ..$ condition    : tibble [1 × 1] (S3: tbl_df/tbl/data.frame)
##   .. ..$ N: num 30
##   ..$ results      :'data.frame':    1000 obs. of  1 variable:
##   .. ..$ p: num [1:1000] 0.00466 0.32682 0.71914 0.01076 0.3534 ...
##   ..$ errors       : 'table' int[0 (1d)] 
##   .. ..- attr(*, "dimnames")=List of 1
##   .. .. ..$ : NULL
##   ..$ error_seeds  : NULL
##   ..$ warnings     : 'table' int[0 (1d)] 
##   .. ..- attr(*, "dimnames")=List of 1
##   .. .. ..$ warnings: NULL
##   ..$ warning_seeds: NULL

Should the need arise to remove the results directory then the SimClean() function is the easiest way to remove all unwanted files and directories.

SimClean(results = TRUE)

Recommendations

My general recommendation when running simulations is to supply a filename = 'some_simulation_name' when your simulation is finally ready for run time (particularly for simulations which take a long time to finish), and to leave the default save = TRUE and store_results = TRUE to track any temporary files in the event of unexpected crashes and to store the results objects for future inspection (should the need arise). As the aggregation of the simulation results is often what you are interested in then this approach will ensure that the results are stored in a succinct manner for later analyses. However, if RAM is suspected to be an issue as the simulation progresses then using save_results = TRUE is strongly recommended to avoid memory-based storage issues.