Terra Optional Param Choices

[samuel-marsh]

Hi Stephen & Team,

Just wondering if you could provide any guidance on couple of the optional parameters when running workflow on Terra for optimal performance (speed mainly).

Given that it runs on GPU using the workflow I'm wondering of the potential hardware_gpu_type whether you have seen performance difference between the various compatible GPUs available via Terra/GCP?

Also wondering if you have rec for hardware_memory_GB (I assume this means GPU memory?) or any of the other optional parameters that give performance gains in terms of run time.

Thanks very much and love the Terra workflow setup!
Sam

[sjfleming]

Hi @samuel-marsh ,

Yes! My favorite GPU for this is the Tesla T4. It has a lot of memory, and it's cheaper (and marginally faster) than the Tesla K80. So that is currently the default "nvidia-tesla-t4". That being said, I've been thinking about cost, so I've only tried out the K80 and the T4, which have lower cost per hour than other GPUs. I guess theoretically it's possible that some of the more expensive and beefier GPUs might be faster, and then the price difference might not be what I'd expect... that I am not sure of, and I have not tested other GPUs yet.

The parameter hardware_memory_GB is for the CPU memory actually. I should make that more clear. But yes, the GPU memory is completely fixed and dictated by the GPU type. It is not adjustable.

Currently, the only other way to get a speedup would be to try to minimize total_droplets_included. But be a little cautious... if this number is too small, then you might end up missing out on some real cells. The idea for this number is to include all the droplets that you think might possibly contain a cell. That is to say, on the UMI curve, all droplets past this point should be "surely empty". But you don't have to make this number too large. Runtime scales linearly with total_droplets_included.

Soon there will be some code updates coming that will reduce the runtime. I am hoping to get those up within the next month or so.

[sjfleming]

@sjfleming make this more clear in the documentation

[sjfleming]

Closed by #238