Conversation
…ed global queuesize to proteindj
eileen-xue
left a comment
eileen-xue
left a comment
There was a problem hiding this comment.
Some GPU tasks requesting 1–2 days of runtime. If the task doesn't need that long, request less time — dgxa100 and gpuhopper nodes are always busy, so shorter requests mean shorter queue waits.
addressing Wenjings Comments
Mitchob
left a comment
Mitchob
left a comment
There was a problem hiding this comment.
@eileen-xue ready for a second review. should mem for hugemembw be 584 or 585 GB?
|
512GB/14*16=585.14GB. |
| cpus = { params.use_dgxa100 ? 16 : 12 } | ||
| gpus = 1 | ||
| memory = '24.GB' | ||
| time = '24.h' |
There was a problem hiding this comment.
Does this need 24h? I don't have data for this right now.
There was a problem hiding this comment.
This version of bindcraft would run until a certain number of attempts (number_of_final_designs) passed QC filters (ie could run very long - extremely variable depending on target protein). Once SBP-536 is merged, we will change to running for a fixed number of attempts (in combination with number of batches) so we can have tighter control over execution time of each individual process. We can set to a more practical runtime now in anticipation of this change.
| cpus = 12 | ||
| time = '2d' | ||
| time = { 5.h * task.attempt } | ||
| memory = '128 GB' |
There was a problem hiding this comment.
Although the job won't throw an error or fail. But from the job log, it needs > 128GB memory. I suggest setting it to 256GB.
There was a problem hiding this comment.
I think 128GB should be (more than) enough for this one but can increase if the cost is the same.
There was a problem hiding this comment.
128GB and 256GB have the same SU cost.
| clusterOptions = '' | ||
| gpus = 1 | ||
| cpus = 12 | ||
| memory = '24GB' |
There was a problem hiding this comment.
Set the memory to 256GB and set a reasonable time for this?
Or leave it at the moment and set a proper number after my benchmarking runs.
3 participants
updates
cleanup
Cleaned up wisps.config by removing settings that duplicated defaults, the file should now only contain genuine NCI/SBP-specific overrides.
more explicitly I cleaned up:
use_gpu,colabfold_num_recycles,ipsae_pae_cutoff/dist_cutoff,colabfold_uniref30_prefix/envdb_prefix)boltz_ccd,boltz_model,uniref30,colabfold_db) not referenced in the pipeline and were giving warningswithLabel:blocks (identical to base.config)shell,ext.args,singularity.registry/runOptions,nextflow.enable.configProcessNamesValidation, and the manifest {} block, and validation.defaultIgnoreParamsThis should have no functional changes.