[Bioc-devel] What is good convention for package-local BiocParallel param?

From the above example, if I had the BPPARAM argument, it?d also clearly add a lot of code noise:

data %>%
apply_methods(method_list1, BPPARAM = MulticoreParam(stop.on.error=FALSE)) %>%
apply_methods(method_list2, BPPARAM = MulticoreParam(stop.on.error=FALSE)) %>%
apply_methods(method_list3, BPPARAM = MulticoreParam(stop.on.error=FALSE))

The compromise is to have

my_bpparam = MulticoreParam(stop.on.error=FALSE)

data %>%
apply_methods(method_list1, BPPARAM = my_bpparam) %>%
apply_methods(method_list2, BPPARAM = my_bpparam) %>%
apply_methods(method_list3, BPPARAM = my_bpparam)

But really at that point why not just have

set_cellbench_bpparam(MulticoreParam(stop.on.error=FALSE))

data %>%
apply_methods(method_list1) %>%
apply_methods(method_list2) %>%
apply_methods(method_list3)

Which I guess is the point of this mail chain. Is there a reason why I shouldn?t?

=On 12 Jan 2019, at 5:43 pm, Aaron Lun <infinite.monkeys.with.keyboards at gmail.com> wrote:

My current set-up in a variety of packages is that every parallelizable function has a BPPARAM= argument. This makes it explicit about which steps are being parallelized. Requiring users to respecify BPPARAM= in each function isn?t as annoying as you?d think, because not many steps are actually heavy enough to warrant parallelization.

Ideally, I'd have BPPARAM=bpparam() by default, allowing me to both respond to the register()'d bpparam() as well as any custom argument that might be supplied by the user, e.g., if they don't want to change bpparam(). However, for various reasons (discussed in the other SerialParam() thread), the current default is BPPARAM=SerialParam().

To be honest, I've never thought it necessary to have a global package-specific parameter for parallelization as you've done (for scPipe, presumably). The current options - global across all packages with register(), or local to individual functions with BPPARAM= - seem to be satisfactory in the vast majority of cases. At least to me.

And at least for RNGs, if a function from another package is giving greatly different results upon parallelization (excepting some numerical error with changed order of summation), I'd say that's a bug of some sort. That should be fixed on their end, rather than requiring other packages and users to tiptoe around them.

-A

On 10 Jan 2019, at 23:59, Shian Su <su.s at wehi.edu.au> wrote:

Hello Developers,

I?m using BiocParallel for parallelism, and I understand that register() is the recommended method for setting threads. But I am reluctant to ask people to run code for my package which changes how other packages operate, so I figured I?d use local bp params. Recent discussions of RNG has made me worried there may be hidden state gotcha?s I?ve not considered. The current implementation is

set_mypkg_threads <- function(n) {
if (n == 1) {
options(?mypkg.bpparam? = SerialParam())
} else if (n > 1) {
if (.Platform$OS.type == "windows") {
options(?mypkg.bpparam" = SnowParam(nthreads))
} else {
options(?mypkg.bpparam" = MulticoreParam(nthreads))
}
}
}

Then elsewhere in my package I make use of parallelism as follows

bplapply(
BPPARAM = getOption(?mypkg.bpparam?, bpparam()),
?
)

Where getOption() either retrieves my set option or the default value given by bpparam(). So the behaviour is that if users have not registered params for my package specifically then it will take the BiocParallel default, but otherwise it will use my package?s local bpparam.

Also I know that as currently implemented, I preclude cluster parallelism on non-Windows machines. But it?s easy to fix. Just looking for feedback on the approach.

Kind regards,
Shian Su