smartdispatch
smartdispatch copied to clipboard
Add a local cluster option to execute on standard linux or mac machine
Keeping track of a hyperparameter search is much easier when using smart-dispatch. It would be nice to have the ability to launch jobs with the [] list notation locally without a cluster.
I don't see yet how to achieve this with at or crontab. There are some ways to do this with a simple bash script like this one: https://stackoverflow.com/questions/6441509/how-to-write-a-process-pool-bash-shell
More options:
xargs
or make
http://coldattic.info/shvedsky/pro/blogs/a-foo-walks-into-a-bar/posts/7
at
and batch
: https://access.redhat.com/documentation/en-US/Red_Hat_Enterprise_Linux/3/html/System_Administration_Guide/s1-autotasks-at-batch.html
I think make
is a good solution. It has a thread pool built in. Here is an example job makefile.
default: job1.log job2.log job3.log
job1.log:
sh job.sh 1 > job1.log
job2.log:
sh job.sh 2 > job2.log
job3.log:
sh job.sh 3 > job3.log
- You can launch make it make -j 2 to have thread pool of size 2.
- It depends on the log files, so if a job has run it will not run again if you run the makefile again.
- The makefile can be generated as the submission script for a cluster.
- Make can be run using
at
so it is a background process on the machine and out of interactive mode.
@mgermain Does the Trello link mean you did some work on your side? We are considering this issue for a CCW ticket.
Just removed the trello link. It was not related. No work has been done for local a local queue and I think there are more pressing issues even if this one would be relly nice to have.
Which issue would you consider more important? A student was asigned the slurm backend ticket. I though he could add support for graham and cedar afterwards, but meanwhile I would have given this ticket or one you find more important to another student.
There are a lot of almost finish, small or low hanging fruit that we could finish before adding another backend which is a much more involved task.
ToFinish/Review/Merge PR
#156 - Add job dependency to autoresume feature #157 - Don't use -V #113 - Write Documentation #42 - Add support for memory
Quick&Small
#140 - Add Coloss support #96 - When msub failed to launch a job we should have some feedback
Important and overdue
#108 - Support custom queues #62 - Add a ~/.smartdispatch config folder #7 - Add more types of folded arguments