search

nanoporetech / dorado

Oxford Nanopore's Basecaller
https://nanoporetech.com/
Other
439 stars 53 forks source link

Increased system RAM usage with dorado 0.7.1 & v5.0.0 models #893

Open samuelmontgomery opened 1 week ago

samuelmontgomery commented 1 week ago

Hi,

I am seeing a much larger use of system RAM when using the v5.0.0 models with dorado 0.7.1 I have a PC with i7 13700K, 64GB of RAM and an RTX4090, as per the recommended specs for the P2 Solo Previously, basecalling would utilise ~20Gb of RAM in addition to the GPU VRAM, but with the new models all 64GB of system ram is being utilised as well

Is this something that will improve with model optimisation? Additionally, is this likely to be a bottleneck for basecalling outside of the GPU?

Thanks - Samuel

malton-ont commented 1 week ago

Hi @samuelmontgomery,

Can you tell us a little more about the command you are running? From a quick test here I'm seeing about 20GB RAM used for basecalling with the sup@v5.0.0 model, which is in line with your expectations. I'm just using:

dorado basecaller sup@v5 pod5s > calls.bam

Is it just changing the models that makes this difference for you? i.e. if you were to use sup@v4.3.0 instead (but still with dorado 0.7.1), does that also have anomalous RAM usage?

samuelmontgomery commented 1 week ago

Thanks - I am running dorado basecaller sup --kit-name SQK-NBD114-96 pod5 > calls.bam using dorado 0.7.1 on Windows

At the moment it's using 23.9GB of VRAM, and 63.4GB of RAM - weirdly when I look at processes in Task Manager, dorado is only using ~2GB of RAM and I cannot see where else it is being used, all other running processes make up ~1GB of RAM

I will have to get back to you later this week with a test dataset - basecalling a rather large P2 run!

On Tue, Jun 18, 2024 at 5:49 PM malton-ont @.***> wrote:

Hi @samuelmontgomery https://github.com/samuelmontgomery,

Can you tell us a little more about the command you are running? From a quick test here I'm seeing about 20GB RAM used for basecalling with the @.*** model, which is in line with your expectations. I'm just using:

dorado basecaller @.*** pod5s > calls.bam

Is it just changing the models that makes this difference for you? i.e. if you were to use @.*** instead (but still with dorado 0.7.1), does that also have anomalous RAM usage?

— Reply to this email directly, view it on GitHub https://github.com/nanoporetech/dorado/issues/893#issuecomment-2175680530, or unsubscribe https://github.com/notifications/unsubscribe-auth/AMKWO6G7JKJK7HOZFY455D3ZH77CHAVCNFSM6AAAAABJPKRZXCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDCNZVGY4DANJTGA . You are receiving this because you were mentioned.Message ID: @.***>