You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi ! so i am trying to model surface spots of a star using starry process (Ensemble Analysis to be precise).For posterior analysis i am using plain vanilla emcee( as recommended in starry_process documentation). whenever i initiate the MCMC with required number of walkers (in my case 30) and steps(in my case 5000) the code runs across all the cores available in the cpu. I have tried to use the threads command of emcee.EnsembleSampler by assigning numerical values to them. Irrespective of me assigning thread value, the code keeps on running across all the available cores.
Surprisingly, when i ran the code for the first time on the system it occupied a single core. While trying to figure out if the speed of the code can be enhanced I ran 'threads' and even then for threads=10 the code ran over all the cores. Since then i am unable to control the number of cores across which we want the code to run
since the time consumption while running across all the cores and single cores is similar we plan on to run multiple ensemble across various single core to fasten our process.
Expected behavior:
it is expected that the code runs over specified no of cores
General information:
linux fedora 5.19.8
Problem description:
Hi ! so i am trying to model surface spots of a star using starry process (Ensemble Analysis to be precise).For posterior analysis i am using plain vanilla emcee( as recommended in starry_process documentation). whenever i initiate the MCMC with required number of walkers (in my case 30) and steps(in my case 5000) the code runs across all the cores available in the cpu. I have tried to use the
threads
command ofemcee.EnsembleSampler
by assigning numerical values to them. Irrespective of me assigning thread value, the code keeps on running across all the available cores.Surprisingly, when i ran the code for the first time on the system it occupied a single core. While trying to figure out if the speed of the code can be enhanced I ran 'threads' and even then for
threads=10
the code ran over all the cores. Since then i am unable to control the number of cores across which we want the code to runsince the time consumption while running across all the cores and single cores is similar we plan on to run multiple ensemble across various single core to fasten our process.
Expected behavior:
it is expected that the code runs over specified no of cores
Actual behavior:
it is running across all the cores
What have you tried so far?:
i have tried
i have tried switching systems of different cores
Minimal example:
The text was updated successfully, but these errors were encountered: