search

nanoporetech / dorado

Oxford Nanopore's Basecaller
https://nanoporetech.com/
Other
488 stars 59 forks source link

2x slowdown with modified RNA basecalling #644

Closed lpryszcz closed 6 months ago

lpryszcz commented 7 months ago

Hi, dorado v0.53 becomes slow with official m6A_DRACH model --modified-bases-models rna004_130bps_sup@v3.0.1_m6A_DRACH@v1 and uses heavily CPU (load over 10 on 16-core CPU) instead of GPU!

[▉                             ] 2% [04h:18m:18s<05d:19h:11m:55s]

Running the same model rna004_130bps_sup@v3.0.1 without m6A_DRACH detection is ~2x faster and uses less CPU (load 3). 

[█▌                            ] 5% [04h:09m:03s<03d:06h:52m:05s]

So the difference seems to be new modification calling algorithm using CPU instead of GPU.

I/O isn't a problem - reading pod5 from local disk, data is hardly access, low GPU use (2x RTX 3080 Ti). Could this be improved somehow?

My command lines: 

~/src/dorado-0.5.3-linux-x64/bin/dorado basecaller rna004_130bps_sup@v3.0.1 --modified-bases-models rna004_130bps_sup@v3.0.1_m6A_DRACH@v1 --estimate-poly-a --emit-sam --emit-moves --reference $ref $f
~/src/dorado-0.5.3-linux-x64/bin/dorado basecaller rna004_130bps_sup@v3.0.1--estimate-poly-a --emit-sam --emit-moves --reference $ref  $f

It's similar to #343 but for RNA, using pod5 and with most recent dorado version.

HalfPhoton commented 6 months ago

Hi @lpryszcz, sorry this reply fell off my radar. We do expect a significant slowdown when adding mods basecalling and as you noticed the mods models are CPU intensive. We're working hard on this and you can expect to see performance improvements in future releases.