Dear All,

My package will use BiocParallel for permutation test.  To make permutations reproducible, I read some discussions online (https://stat.ethz.ch/pipermail/bioc-devel/2015-June/007583.html, https://stackoverflow.com/questions/30610375/how-to-run-permutations-using-mclapply-in-a-reproducible-way-regardless-of-numbe). There are two easy solutions:

(1) Generate the permuted sample indexes before parallel computations. In this way, set.seed() can be used outside of my R function.

(2) Generate a vector of seeds (provided by users through argument of my R function) and use them by set.seed() in each parallel computation.

My permutation scheme is a little complex, so storing sample indexes needs a huge storage space. I have to use the second solution. However, set.seed() will cause warning in BiocCheck(). Could anybody give me suggestions?

Thanks a lot for your help!

Best,

Lulu

This question is more appropriate for the Bioc-devel mailing list. Please direct future correspondence there.

In the meantime, here's some food for thought. I often run into this problem in several of my packages containing with heavy-duty randomized algorithms, e.g., correlation functions in scran, emptyDrops in DropletUtils. Fortunately for me, all of the functions that do enough computationally intensive work to warrant parallelization have also been re-written in C++. Thus, in the serial phase of each function, I use R's PRNG to generate random integers, one per computational job (e.g., per sample or per iteration). In each job, the sampled integer is used to seed std::mt19937 in each worker for random number generation at the C++ level. No matter how the jobs are split up between workers, I will always obtain the same series of random numbers. This ensures that I obtain the same results regardless of the BiocParallelParam object used. It also leaves R's .Rand.seed in the same state regardless of the parallelization scheme, which ensures that the output of downstream random functions are not dependent on whether I parallelized an upstream function.