Fix mem sort for HISAT2 output

[apeltzer]

Copying in the conversation from nf-core slack:

So we keep seeing some issues with the RNAseq workflow when sorting the hisat2 output - memory requested on the server seems to be higher than what is provided, thus the jobs get killed by the scheduler...
e.g. something like Excceed job memory limit 63348736 > 82914560
which happens here

rnaseq/main.nf

Line 676 in e837637

-@ ${task.cpus} $avail_mem \\

I'd suggest dropping this memory request, as I never really saw issues without it in other pipelines sorting BAM files...
Sarek e.g. does this somewhat similar thing to set it permanently to 2G https://github.com/SciLifeLab/Sarek/blob/88921e6944604d077afd8a5bc0a89b8e93937313/main.nf#L174
Suggestions ?

[ewels]

Wow, that's like 30% over... Are you sure that the final script is actually using the -m flag? Note that it can be empty depending on the config.

We should probably at least leave some overhead from that calculation though at the very least.

[apeltzer]

Hm, what I thought when looking at the command is that the -m <XYZ> flag isn't used at the moment. The command simply states the $avail_mem instead of using the -m $avail_mem value, which might cause some trouble too for big bam files? Maybe that is the real reason?

I'd also be happy to compute the avail_mem as is and then deduct something as you mentioned it - let me know and I'll adapt it, maybe trying first to just incorporate the -m switch ?

[ewels]

I think the -m flag is inside that variable, no? It's intentional - so that the -m flag isn't left empty if we don't know what memory the process has.

[apeltzer]

It is indeed - I changed to using - 100000000 as in hisat, star and other memory intensive processes to keep hopefully within the limits 👍 Sorry my bad, I didn't see it...

[ewels]

Nice! Is it possible for you to test this before merging to see if it fixes the problem? Or should we just merge?

[apeltzer]

I can let Silvia test tomorrow 👍

[apeltzer]

We're testing at the moment - will keep you posted on updates

[apeltzer]

Local testing revealed that this seems to work fine even for weird samples that were not running beforehand 👍

[ewels]

Nice! Could you just add in some extra code that checks that this isn't a negative number / stupidly small? If so then it can just be a blank string and left unspecified again. I'm just thinking that on some small (eg. testing) envs, there may be under 6GB memory given to this process.

[apeltzer]

Done :-)