Inquiry Regarding Slow Performance with Dorado Basecaller

[xiangpingyu]

Dear developers,

The size of one dataset is about 80.0GB, and it's stored on a local disk (when running, GPU memory = 34.3/55.8 GB) the following is the CPU and GPU info:

• CPU： 13th Gen Intel® core i9-13900K, Cores=24, logical processors=32

• GPU: NVIDIA GeForce RTX 4090, CUDA cores=16384

This is the command I am using

dorado basecaller sup,6mA --no-trim --recursive ./ > m6A.bam

I am unsure when the process will complete. Is there anything I need to modify?

Thank you!

[HalfPhoton]

HI @xiangpingyu,

• Is there any data in the output file m6A.bam?
• Can you share what you see if you add verbose logging output with -vv?
• Are you using pod5 or fas5?

Kind regards, Rich

[iiSeymour]

@xiangpingyu the eta for this job is just over 1 hour. Are you saying it's making no progress?

[xiangpingyu]

@HalfPhoton

• I use the pod5.

• the following as for your reference, with -vv

• Running for ~5 hours, it is only 1% processed $ dorado basecaller [email protected],6mA@v2 ./ -x "cuda:all" --no-trim -r --kit-name SQK-NBD114-24 > 6ma_calls.bam

[xiangpingyu]

@iiSeymour it's my first time to run dorado in the system. I'm unsure that how long it will finish. Now, about five hours, the task only came up to ~1%, based on the previous information.

[tijyojwad]

Hi @xiangpingyu can you run nvidia-smi and report what the GPU utilization numbers look like? You can also install nvtop utility which will show a nice utilization graph over time which can help us understand how well basecalling is utilizing the GPU.

Also, what is the expected read length distribution of your data?

[xiangpingyu]

Hi @xiangpingyu can you run nvidia-smi and report what the GPU utilization numbers look like? You can also install nvtop utility which will show a nice utilization graph over time which can help us understand how well basecalling is utilizing the GPU.

Also, what is the expected read length distribution of your data?

@tijyojwad the following is the output of "nvidia-smi" and the performance of GPU. The expected read length is in the range of 2.5k ~ 6kb.

[jemoore16]

@xiangpingyu Did it take 2 weeks for the base-calling to complete?

[HalfPhoton]

@xiangpingyu Do you have any updates on this issue?

[axelsose]

Hi, I am also experiencing very slow performance when using Dorado basecaller from command line with the modified bases models (see command in the picture below):

The BAM file is 17100 MB at the moment, with only ~5% of the data basecalled after more than one day.... The total dataset is 2.2T obtained with P2 solo.

Find also a screenshot of the nvidia-smi

Any suggestion to speed up the process? When I do the basecalling from MinKnow it is way faster, but then I can only choose one model for the modified bases, and in this case I need to detect 6mA, 4mC, and 5mC all together.

Many thanks in advance! Axel

[HalfPhoton]

The v5 sup models are are under active development to improve basecalling performance. They use a new architecture which has yet to be optimised fully. Combining this and detecting 3 mods is a heavy computational load and as such the basecalling performance is slow.

Minknow is much quicker because it will be using v4.3 models which uses the old architecture which has had a large amount of effort put into its performance.