Memory usage for ultra long read alignment

[Daniel Liu]

@RagnarGrootKoerkamp and I have observed the memory usage of BiWFA increasing significantly to >10GB when aligning >500kbp ONT reads with linear and affine costs (eg., sub = 1, open = 1, extend = 1). We are currently using Rust bindings here (using cmake to build with flags recommended by the WFA2-lib readme) and we have observed this memory usage issue on both the latest main branch and a commit from around 1 year ago. We have reproduced this on both MacOS and Linux.

The Rust wrapper is quite thin so we think it's unlikely that its the cause of the issue. Any idea what's wrong?

[Ragnar Groot Koerkamp]

Note: These are the same reads as the ONT UL in figure 2 of the BiWFA paper, so this is likely some regression with the code at some point. With unit costs (sub=extend=1, open=0), memory usage is very small as expected.

[Ragnar Groot Koerkamp]

I ran it through heaptrack, and unsurprisingly, all the memory comes from wavefront_allocate via mm_allocator_segment_new

[Santiago Marco-Sola]

Hi,

Thanks for the report, guys. I will have a look. Note that, if you see the "unialign" functions on the profile, it means that you are using regular WFA. The BiWFA uses all the "bialign" modules.

In any case, I will have a look.

Thanks,

[Daniel Liu]

That prompted me to take another look at the bindings and I see the issue: https://github.com/pairwise-alignment/rust-wfa2/blob/main/src/aligner.rs#L321
Some of the parameters (memory model and alignment scope) are being overwritten by this default() statement.

[Daniel Liu]

Yeah that seems to be it. Sorry for bothering you!

[Santiago Marco-Sola]

Anytime, Daniel. Let me know how it goes.