An open API service indexing awesome lists of open source software.

https://github.com/kylebaron/mrgsim.parallel

Parallel simulation with mrgsolve and futures
https://github.com/kylebaron/mrgsim.parallel

future mrgsolve parallelization

Last synced: about 2 months ago
JSON representation

Parallel simulation with mrgsolve and futures

Awesome Lists containing this project

README

          

---
title: ""
output: github_document
---

```{r,setup,include=FALSE}
knitr::opts_chunk$set(comment = '.', message=FALSE, warning = FALSE,
fig.path="man/figures/README-")
```

# mrgsim.parallel

## Overview
mrgsolve.parallel facilitates parallel simulation with
mrgsolve in R. The future and parallel packages provide the parallelization.

There are 2 main workflows:

1. Split a `data_set` into chunks by ID, simulate the chunks in parallel, then
assemble the results back to a single data frame.
1. Split an `idata_set` (individual-level parameters) into chunks by row,
simulate the chunks in parallel, then assemble the results back to a single
data frame.

The nature of the parallel backend requires some overhead to get the
parallel simulation done. So, it will take a reasonably-sized job to see
a speed increase and small jobs will likely take *longer* with parallelization.
But jobs taking more than a handful of seconds could benefit from this type
of parallelization.

```{r,include = FALSE}
options(mrgsolve.soloc = "build")
```

## Backend

```{r}
library(dplyr)

library(future)

library(mrgsim.parallel)

options(future.fork.enable = TRUE, parallelly.fork.enable = TRUE, mc.cores = 4L)
```
## First workflow: split and simulate a data set

```{r}
mod <- modlib("pk2cmt", end = 168*8, delta = 1)

data <- expand.ev(amt = 100*seq(1,2000), ii = 24, addl = 27*2+2)

data <- mutate(data, CL = runif(n(), 0.7, 1.3))

head(data)

dim(data)
```

We can simulate in parallel with the future package or the parallel package like this:
```{r}
plan(multisession, workers = 4L)
system.time(ans1 <- future_mrgsim_d(mod, data, nchunk = 4L))

plan(multicore, workers = 4L)
system.time(ans1b <- future_mrgsim_d(mod, data, nchunk = 4L))

system.time(ans2 <- mc_mrgsim_d(mod, data, nchunk = 4L))
```

To compare an identical simulation done without parallelization
```{r}
system.time(ans3 <- mrgsim_d(mod,data))
```

```{r}
identical(ans2,as.data.frame(ans3))
```

## Second workflow: split and simulate a batch of parameters

Backend and the model
```{r}
plan(multisession, workers = 6)

mod <- modlib("pk1cmt", end = 168*4, delta = 1)
```

For this workflow, we have a set of parameters (`idata`) along with an
event object that gets applied to all of the parameters
```{r}
idata <- tibble(CL = runif(4000, 0.5, 1.5), ID = seq_along(CL))

head(idata)
```

```{r}
dose <- ev(amt = 100, ii = 24, addl = 27)

dose
```

Run it in parallel
```{r}
system.time(ans1 <- mc_mrgsim_ei(mod, dose, idata, nchunk = 6))
```

And without parallelization

```{r}
system.time(ans2 <- mrgsim_ei(mod, dose, idata, output = "df"))

identical(ans1,ans2)
```

## Utility functions

You can access the chunking functions for your own parallel workflows

```{r}
dose <- ev_seq(ev(amt = 100), ev(amt = 50, ii = 12, addl = 2))
dose <- ev_rep(dose, 1:5)

dose

chunk_by_id(dose, nchunk = 2)
```

See also: `chunk_by_row`

## Do a dry run to check the overhead of parallelization

```{r}
plan(transparent)
system.time(x <- fu_mrgsim_d(mod, data, nchunk = 8, .dry = TRUE))

plan(multisession, workers = 8L)
system.time(x <- fu_mrgsim_d(mod, data, nchunk = 8, .dry = TRUE))

```

## Pass a function to post process on the worker

First check the range of times from the previous example

```{r}
summary(ans1$time)
```

The post-processing function has arguments the simulated data and the
model object
```{r}
post <- function(sims, mod) {
filter(sims, time > 600)
}

dose <- ev(amt = 100, ii = 24, addl = 27)

ans3 <- mc_mrgsim_ei(mod, dose, idata, nchunk = 6, .p = post)
```

```{r}
summary(ans3$time)

```

The main use case here is to summarize or some how decrease the volume of data
before returning the combined simulations. In case memory is able to handle
the simulation volume, this post-processing could be done on the combined
data as well.


## More info

See [inst/docs/stories.md (on GitHub only)](inst/docs/stories.md) for more details.