Skip to contents

Efficiency of tumopp

Source: vignettes/benchmark.Rmd
benchmark.Rmd

Run simulation with benchmark information 

library(tidyverse)
library(tumopp)

.const = list(D = 3, C = "hex", k = 10, N = 2 ** 17, benchmark = TRUE)
.alt = list(
  L = c("const", "linear"),
  P = c("random", "mindrag"),
  d = c(0, 0.2),
  m = c(0, 2)
)
args_tbl = make_args(alt = .alt, const = .const) |> print()

results = tumopp:::system_tumopp(args_tbl) |> print()

unnested = results |>
  dplyr::select(!c(population, seed)) |>
  dplyr::mutate(outdir = fs::path_file(outdir)) |>
  dplyr::select(where(\(x) dplyr::n_distinct(x) > 1L)) |>
  tidyr::unnest(benchmark) |>
  print()

df_benchmark = unnested |>
  dplyr::distinct(outdir, size, .keep_all = TRUE) |>
  dplyr::mutate(
    local = factor(local, levels = .alt[["L"]]),
    path = factor(path, levels = .alt[["P"]]),
    size = size / 1000,      # to K
    memory = maxrss / 1024,  # K to M
    utime = utime / 1000,    # ms to sec
    stime = stime / 1000,    # ms to sec
    maxrss = NULL
  ) |>
  print()

# usethis::use_data(df_benchmark, internal = TRUE, overwrite = TRUE)

Memory

  • Memory usage increases linearly with the increasing number of cells because tumopp is an individual-based model of tumor cells.
  • Memory is also used to trace back cell lineages and write them to a text file after growth simulation.
  • Cell death makes cell lineages longer given the same number of extant cells.
  • The effect of cell death is more pronounced when the population average of birth rate is decreased with local competition (-Llinear).
ggplot(tumopp:::df_benchmark) +
  aes(size, memory) +
  geom_line(linewidth = 1, color = "#337ab7") +
  facet_grid(vars(local, path), vars(delta0, rho0), label = label_both) +
  labs(x = "Size (K cells)", y = "Memory (MB)")

Time

  • The larger a population gets, the longer it takes
    • (-Pmindrag) to search for nearest empty sites from dividing cells,
    • (-Lconst) to push out other cells,
    • (-Llinear) to skip event queues of dormant (surrounded) cells.
  • The combination of -Lconst -Pmindrag is slower than the others because it involves both search and push from interior cells.
  • The effects of cell death and migration are smaller than those of local competition (-L) and placement of daughter cells(-P).
ggplot(tumopp:::df_benchmark) +
  aes(size, utime + stime) +
  geom_line(linewidth = 1, color = "#337ab7") +
  facet_grid(vars(local, path), vars(delta0, rho0), label = label_both) +
  labs(x = "Size (K cells)", y = "Time (sec)")