Reduce memory usage of facets

[natir]

Hi,

In my laboratory we use facets on many whole genome human dataset, on this data facets have a huge memory usage, approximately 150 Gib.

The purpose of this PR is to try to reduce facets memory usage, for this I replace some classic R data.frame by tidyverse tibble data-structure, I also use tydiverse pipe syntaxe to perform some operation on this tibble.

With all this change, I divide memory usage by 2.

On my test dataset result is same between my PR and version v0.6.1, but maybe I miss some stuff.

I'm not a good R developer, maybe I include some stupid mistake, so if you want just take the idea of my change and rewrite it please do it.

Thank

[veseshan]

Can you give me some breakdown of where this memory explosion occurs. My back of the envelope calculation says

R:> x = rnorm(12e6) # one locus every 250 bases across 3000 Megabase
R:> format(object.size(x), units="Mb")
[1] "91.6 Mb"

The jointseg data frame has 16 columns but even that wouldn't translate to 150Gib memory use.

Have you tried using the readSnpMatrixDT.R in path/facets/extRfns/ to read in the data?

Thanks

[natir]

With v0.6.1 the memory peak is during file reading, use readSnpMatrixDT.R like my change solve this issue.

But another peak occur during preProcSample I assume, it's more specifically in procSnps (some duplication, column creation, calling of Fortran code and filtration not run in place).

With v0.6.1 and readSnpMatrixDT.R memory usage is 85Gib, my version use 70Gib.

[veseshan]

Can you tell me how big is the pileup matrix i.e. how many loci? And how many end up in jointseg? Thanks.

[natir]

The pileup matrix contains 546,700,164 loci.

To evaluate number of jointseg I consider $jointseg in output produce by procSample, I get 5,583,831 jointseg.

[veseshan]

Given that the whole genome is around 3 Gigabase, the pileup seems to have a locus every 6 bases. That is a lot of redundant data as they will be highly serially correlated. You can DM me if you want to talk about this further.

I will look into how your code can be used to reduce the memory use of procSnps.

Thanks