dorado polish - RAM of 768 GB not sufficient

[colindaven]

New issue checks

• I have read the Dorado Documentation.
• I did not find an existing issue.

Dorado version

1.3.1

Dorado subcommand

Polish

The issue

Hello,

I'm trying to get dorado polish working properly. Having given up on gpu (likely a nextflow submission problem) and slurm (out of control RAM use) I am now trying to debug the polishing RAM requirements on a local server. I've read similary issues like #1524

Server specs - 64 cores, 768 GB RAM.

I have had over 30 attempts with various combinations of threads (8-24), batch sizes (0, 128) and infer-threads (2-24) with no success. This is for a medium sized 800 MB plant genome. With medaka I polish genomes of this size and larger regularly.

Result is always oom kill / out of RAM.

Command:
[2026-01-22 12:10:30.458] [info] Running: "polish" "-x" "cpu" "--batchsize" "128" "--ignore-read-groups" "--threads" "8" "--infer-threads" "2" "dorado_align_bams.bam" "contigs-unpolished.fasta"

Surely this server should be big enough to handle dorado polish ? No other jobs are running. I have bigger servers but people would be unhappy to see me using 2 TB of RAM but only 8 cores. Or does the RAM not increase with more cores ?

Options going forward

• drastically reduce coverage (perhaps 20-30x is sufficient, I often have 50-80x)
• use medaka instead
• optimize settings?

Thanks
Colin

System specifications

linux, nextflow, slurm or local

[colindaven]

One further point.

Note that I have consistently got 200% so 2 threads running on cpu, yet I have requested --threads 8. I will remove the --infer-threads 2 part - as I believe that is only supposed to affect gpu resource allocation. This seems to be a bug though if it affects pure cpu processes ?

200.7 6.9 2274:21 dorado polish -x cpu --batchsize 128 --ignore-read-groups --threads 8 --infer-threads 2 dorado_align_bams.bam 2651947-ZR_2xxxx

[colindaven]

I restarted and dorado polish ran through with 32 threads set by --threads 32 but only using 2 cores (as evidenced by 200% median cpu usage by nexflows report).

This took 41 hours for this circa 700 MB plant genome using 560 GB of RAM. The coverage was reduced to about 30X prior to polishing.

Is there a reason why it is only using 2 cpu cores ? Is this intended ?

Thanks

[malton-ont]

Hi @colindaven,

As you're using CPU inference, I suspect that the process is being limited by the inference speed. The default value for --infer-threads is 2. You could try increasing this value.