Running scripts like the stabilization and cropping were causing a slowdown in the I/O for other users @talonchandler @ieivanov
After some debugging, I believe I've nailed down the issue. This issue can be replicated in two ways with datasets that have lots of positions and timepoints and/or channels.
Most slurm scripts parallize over T and C and submit individual jobs for each positions. Keep in mind that since we parallelize over these two dimensions each multiprocessing pool will spawn n number of subprocesses. So, if all positions are allocated, then each jobs will spawn its respective number of child processes causing lots of I/O calls.
Make a slurm script with more CPUs, memory and simultaneous number of subprocesses. If we do this we run into the same issue as we will have multiple jobs running in parallel and I/O calls which reduce our overall throughput.
Proposed solutions:
Make sure that the total amount of jobs per datasets is below a TBD threshold. The combination of T and C simultaneous processes and number of positions should be less than this threshold. I need to find a sweet spot for this.
Modify the slurmscripts so that we run the positions in batches. This can be done by creating dependencies between the jobs as follows:
batch_size = 3
register_jobs = []
for i in range(0, len(input_position_dirpaths), batch_size):
chunk_input_paths = input_position_dirpaths[i : i + batch_size]
if i == 0:
register_jobs = [
submit_function(
register_func,
slurm_params=params,
input_data_path=in_path,
output_path=out_path,
)
for in_path in chunk_input_paths
]
else:
register_jobs = [
submit_function(
register_func,
slurm_params=params,
input_data_path=in_path,
output_path=out_path,
dependencies=register_jobs,
)
for in_path in chunk_input_paths
Problem
Running scripts like the stabilization and cropping were causing a slowdown in the I/O for other users @talonchandler @ieivanov
After some debugging, I believe I've nailed down the issue. This issue can be replicated in two ways with datasets that have lots of positions and timepoints and/or channels.
nnumber of subprocesses. So, if all positions are allocated, then each jobs will spawn its respective number of child processes causing lots of I/O calls.Proposed solutions:
for i in range(0, len(input_position_dirpaths), batch_size): chunk_input_paths = input_position_dirpaths[i : i + batch_size] if i == 0: register_jobs = [ submit_function( register_func, slurm_params=params, input_data_path=in_path, output_path=out_path, ) for in_path in chunk_input_paths ] else: register_jobs = [ submit_function( register_func, slurm_params=params, input_data_path=in_path, output_path=out_path, dependencies=register_jobs, ) for in_path in chunk_input_paths