Adding %doFuture% parallel framework for the sim_gs_n.R code and upda…

[cmansch]

Implementation for sim_gs_n() to have a parallel adaptor similar to sim_fixed_n() (xref: #110).
A seed is now set prior to the function call instead of passed as an argument.
Function now uses %doFuture% to run iterations of n_sim in parallel.

Addresses open issue (xref: #205).

Sequential backend still available (and default) to avoid any delays in development.

Please update "Remove seed argument from sim_gs_n() in (xref: Merck/gsDesign2#288) as part of the implementation.

New pull request created after (xref: #229) merged.
See old PR as (xref: #230)

[jdblischak]

We should add test cases for both the doFuture and doParallel backends. I'm pretty sure the GitHub Actions machines have 2 cores available. And if we're worried about CRAN, we can always add skip_on_cran() for these tests of the parallelization method

[nanxstats]

I agree on adding tests. CRAN enforces the max 2 core policy strictly so to avoid any potential trouble, always using skip_on_cran() is a good idea, just in case someone accidentally added something that used all cores.

[jdblischak]

Might be worth a brief mention in the vignette on parallelization that now sim_gs_n() supports the same parallelization behavior as sim_fixed_n()

[jdblischak]

Any idea why we are now getting the error object 'results' not found? We didn't edit the vignettes, so I assume this is due to the change in the error handling behavior of sim_fixed_n().

Confirmed. When I revert sim_fixed_n() to use .errorhandling = "pass", I can build the vignettes locally.

[jdblischak]

I confirmed the error handling behavior of sim_gs_n(). It immediately fails when running in parallel and returns a single error, similar to sequential processing. I used Example 8 (multiple tests per cut), which isn't supported yet, and thus should fail.

library("future")
library("simtrial")

source("tests/testthat/helper-sim_gs_n.R")

ia1_test <- list(
 test1 = create_test(wlr, weight = fh(rho = 0, gamma = 0)),
 test2 = create_test(wlr, weight = fh(rho = 0, gamma = 0.5)),
 test3 = create_test(rmst, tau = 20)
)
ia2_test <- list(
 test1 = create_test(wlr, weight = fh(rho = 0, gamma = 0)),
 test2 = create_test(wlr, weight = early_zero(6))
)
fa_test <- list(
 test1 = create_test(wlr, weight = fh(rho = 0, gamma = 0)),
 test3 = create_test(milestone, ms_time = 20)
)

sim_gs_n(
 n_sim = 3,
 sample_size = 400,
 enroll_rate = test_enroll_rate(),
 fail_rate = test_fail_rate(),
 test = list(ia1 = ia1_test, ia2 = ia2_test, fa = fa_test),
 cut = test_cutting()
)
## Backend uses sequential processing.
## Loading required package: foreach
## Error: attempt to apply non-function

plan("multisession")

sim_gs_n(
  n_sim = 3,
  sample_size = 400,
  enroll_rate = test_enroll_rate(),
  fail_rate = test_fail_rate(),
  test = list(ia1 = ia1_test, ia2 = ia2_test, fa = fa_test),
  cut = test_cutting()
)
## Using 16 cores with backend cluster
## Error: attempt to apply non-function

[jdblischak]

Looking good. Still needs some examples and tests of sim_gs_n() running in parallel.

Also, given that updating sim_fixed_n() is causing problems, maybe we should separate the sim_fixed_n() updates into a separate PR.

[cmansch]

I have reverted the sim_fixed_n() error handling and created a new issue.

Updating with a parallel example will come soon.

[cmansch]

Commit added with an example using parallel processing that will be skipped by CRAN similar to in sim_fixed_n(). Similarly, a parallel test has been added that performs the logrank test with a multisession backend. The other future backends do not work on all machines and are therefore omitted from tests.

[jdblischak]

Great work @cmansch! I just had one minor comment about the test format, but I don't want to delay merging this PR. We can always address it in a future PR if desired

[jdblischak]

Instead of hard-coding the expected output, you could instead set the seed again and run the exact same call to sim_gs_n() in sequential mode. That would be more succinct and wouldn't require any manual updating if we update the output of sim_gs_n()

[LittleBeannie]

Thanks @cmansch for leading this pull request, and I extend my thanks to everyone who has contributed to this exceptional effort. I will proceed with merging it.