Meta tags:

Headings (most frequently used words):

dask, jobqueue, contents, dynamic, clusters, batch, runners, more, details,

Text of the page (most frequently used words):
dask (28), the (21), and (13), clusters (12), #jobqueue (11), workers (11), client (11), #cluster (11), runners (10), dask_jobqueue (10), more (8), batch (8), dynamic (8), this (8), can (7), are (6), job (6), import (6), workload (6), how (5), use (5), all (5), jobs (5), distributed (5), scheduler (5), api (4), overview (4), configuration (4), works (4), documentation (4), small (4), hpc (4), from (4), slurm (4), where (4), for (4), development (3), details (3), installing (3), your (3), examples (3), talks (3), tutorials (3), processes (3), large (3), many (3), single (3), process (3), slurmrunner (3), srun (3), code (3), during (3), then (3), run (3), slurmcluster (3), contents (2), history (2), guidelines (2), changelog (2), debug (2), help (2), reference (2), writing (2), own (2), advanced (2), tips (2), tricks (2), example (2), deployments (2), configure (2), interactive (2), getting (2), started (2), point (2), learn (2), see (2), runner (2), same (2), start (2), allocation (2), available (2), when (2), starts (2), submitting (2), other (2), workloads (2), generally (2), with (2), benefits (2), commands (2), now (2), these (2), resources (2), array (2), connect (2), remote (2), batch_workload (2), pattern (2), well (2), node (2), submitted (2), queue (2), compute (2), scale (2), times (2), that (2), new (2), dynamic_workload (2), common (2), deploy (2), queuing (2), systems (2), g55b9727 (2), copyright, 2018, anaconda, inc, contributors, team, next, good, entry, know, about, have, wall, time, comprised, one, tends, reliable, than, allocations, guaranteed, which, avoid, oversubscribing, frowned, upon, some, hpcs, typical, colocated, physically, machine, communication, faster, expecially, ucx, only, will, continue, past, boostraps, into, python, prioritised, locality, important, whole, including, within, coordinate, startup, work, together, you, don, need, wait, nodes, before, often, sooner, intaractive, environments, down, users, manually, slowly, few, busy, quiet, gaps, would, otherwise, unused, backfilled, autoscale, progresses, local, tells, call, needs, maximum, minimum, adapt, favourable, submit, environment, means, location, usually, additional, added, dynamically, while, running, there, two, deployment, patterns, has, support, both, project, makes, easy, typically, found, high, performance, supercomputers, academic, research, institutions, easily, like, pbs, moab, sge, lsf, htcondor, pdf, rst, sgecluster, pbscluster, oarcluster, moabcluster, lsfcluster, htcondorcluster, ctrl, back, top, community, ecosystem, skip, main, content,

Text of the page (random words):
dask jobqueue dask jobqueue 0 9 0 14 g55b9727 documentation skip to main content dask distributed dask ml examples ecosystem community back to top ctrl k dask jobqueue 0 9 0 14 g55b9727 documentation getting started installing talks tutorials dynamic clusters overview interactive use how this works configuration configure dask jobqueue example deployments configuration examples advanced tips and tricks clusters api dask_jobqueue htcondorcluster dask_jobqueue lsfcluster dask_jobqueue moabcluster dask_jobqueue oarcluster dask_jobqueue pbscluster dask_jobqueue sgecluster dask_jobqueue slurmcluster batch runners overview writing your own runners runners api dask_jobqueue slurm slurmrunner help reference how to debug changelog development guidelines history rst pdf dask jobqueue contents dynamic clusters batch runners more details dask jobqueue easily deploy dask on job queuing systems like pbs slurm moab sge lsf and htcondor the dask jobqueue project makes it easy to deploy dask on common job queuing systems typically found in high performance supercomputers academic research institutions and other clusters there are two common deployment patterns for dask on hpc dynamic clusters and batch runners and dask jobqueue has support for both dynamic clusters a dynamic cluster is a dask cluster where new workers can be added dynamically while the cluster is running in an hpc environment this generally means that the dask scheduler is run in the same location as the client code usually on a single compute node then workers for the dask cluster are submitted as additional jobs to the job queue scheduler this pattern works well on clusters where it is favourable to submit many small jobs srun n 1 dynamic_workload py dynamic_workload py from dask_jobqueue slurm import slurmcluster cluster slurmcluster cluster adapt minimum 1 maximum 10 tells dask to call srun n 1 when it needs new workers from dask distributed import client client client cluster connect this local process to remote workers import dask array as da x dask commands now use these distributed resources benefits clusters can autoscale as a workload progresses small gaps in the hpc that would be otherwise unused can be backfilled a workload can run slowly with a few workers during busy times and then scale up during quiet times workloads in intaractive environments can scale up and down as users run code manually you don t need to wait for all nodes to be available before your workload starts so jobs often start sooner to learn more see the dynamic cluster documentation batch runners a batch runner is a dask cluster where the whole workload including the client code scheduler and workers are submitted as a single allocation to the job queue scheduler all of the processes within the workload coordinate during startup and then work together to compute the dask workload this pattern works well where large jobs are prioritised and node locality is important srun n 12 python batch_workload py batch_workload py from dask_jobqueue slurm import slurmrunner cluster slurmrunner boostraps all the processes into a client scheduler 10 workers only the client process will continue past this point from dask distributed import client client client cluster connect this client process to remote workers import dask array as da x dask commands now use these distributed resources benefits workers are generally colocated physically on the machine so communication is faster expecially with ucx submitting many small jobs can be frowned upon on some hpcs submitting a single large job is more typical of other hpc workloads all workers are guaranteed to be available when the job starts which can avoid oversubscribing workers clusters comprised of one large allocation tends to be more reliable than many small allocations all processes have the same start and wall time to learn more see the batch runner documentation more details a good entry point to know more about how to use dask jobqueue is talks tutorials getting started installing talks tutorials dynamic clusters overview interactive use how this works configuration configure dask jobqueue example deployments configuration examples advanced tips and tricks clusters api batch runners overview writing your own runners runners api help reference how to debug changelog development guidelines history next installing contents dynamic clusters batch runners more details by dask jobqueue development team copyright 2018 anaconda inc and contributors