JISCMail - CCPEM Archives

Pranav,

It depends on your goals, but if you want to validate subtraction results, conversion between programs, make masks, or obtain initial references, then the simplest thing is to use fewer particles. If you need the highest resolution reconstruction possible, then you have to wait.

That said, you can probably parallelize more aggressively. Since this is an IO bound task, more ranks above the number of available cores can speed things up. Check the actual IO rate using iotop or htop (you can add the IO rates via the setup menus). I usually use 40 MPI ranks for reconstruction on a single workstation with a 6-disk RAID 10. For ~300,000 particles on a ~256px box I see just a couple minutes for backprojection and around 10-15 minutes for the reconstruction. You can find the number of ranks the maximizes the IO rate, or minimizes relion's estimated time required.

You can also play with the Fourier padding and interpolation setting to speed it up, but it could compromise the resolution. Whether these parameters or fewer particles is better is likely data dependent.

Another approach is to import into cisTEM and use "Generate 3D," it's very fast, but of course you have to wait for the import and refinement package creation and use another program.

Best,

-da

On Tue, Sep 4, 2018 at 11:59 AM, Pranav Shah <[log in to unmask]> wrote:

Dear Community,

I am running on a bunch of tests on a set of refined particle stack with different kind of masks, 3D classifying the particles, and reconstructing the volumes from interesting classes without the mask from the command line using relion_reoncstruct_mpi. However, these reconstructions take a really long time since they are using the CPUs and not the GPUs. I was wondering if anyone would know a "hack" to speed up simple reconstruction calculations from the command line. For your reference, the command syntax is as below.

mpirun -np 17 `which relion_reconstruct_mpi` --i input.star --o output.mrc --angpix 1.065 --sym I1 like this

I am using relion 3.0 beta.

Looking forward to your suggestions.

Best,
Pranav

--
Pranav Shah
Postdoctoral Research Fellow.

Hogle Lab
Harvard Medical School
240 Longwood Avenue
Boston, MA 02115
(617) 432-4360 (fax)
(617) 432-3839 (lab)

To unsubscribe from the CCPEM list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=CCPEM&A=1