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
Originally reported by: Dimitry Tegunov (Bitbucket: DTegunov, GitHub: DTegunov)
No matter what I set in the GUI, when I press 'print command', it sets --pool 1 if --preread_images is set. I understand that --pool is meant to reduce the number of disk accesses, and that aspect becomes irrelevant when all images are already in memory. However, doesn't it also provide better load balancing within each MPI process? I'm getting better performance if I set it to 30, despite --preread_images. Maybe the default shouldn't be 1?
Original comment byBjoern Forsberg (Bitbucket: bforsbe, GitHub: bforsbe):
You are correct that load-balancing is improved with increased --pool, but between threads. MPI-load balancing is essentially irrelevant for RELION. If one rank were to entirely halt, all other keep going until all particle images are done. This is why MPI-ranks are in general better to use, but they have a drastic memory-penalty and increased overhead in other aspects, which makes it useful to use a single rank and many threads on each GPU. To load-balance this, --pool is key, yes.
Originally reported by: Dimitry Tegunov (Bitbucket: DTegunov, GitHub: DTegunov)
No matter what I set in the GUI, when I press 'print command', it sets --pool 1 if --preread_images is set. I understand that --pool is meant to reduce the number of disk accesses, and that aspect becomes irrelevant when all images are already in memory. However, doesn't it also provide better load balancing within each MPI process? I'm getting better performance if I set it to 30, despite --preread_images. Maybe the default shouldn't be 1?
The text was updated successfully, but these errors were encountered: