time needed to reconstruct a 100 000 particles volume, and ways to speed it up.
Forums
Hello,
Doing a refinement on a small dataset, I see that the most time consuming step is the reconstruction, as it is done on one single node. For 18 000 particles, 350x350, it takes about 5 hours. For 100 000 particles, I guess that it's going to be ~28 hours. It will cut that time a little bit once I get the multicore to work, but is there parameters that could be set to speed up the process?
If I skip the reconstruction altogether with RELMAG set to -100, is it still doing an iterative refinement? I am not sure to understand how particle parameters are refined in this procedure. (I read your 2007 article, but am not entirely sure of my level of understanding of that part..)
Another way to circumvent the problem would be to split the reconstruction over a few nodes and merge the volumes after. Is that a bad idea, providing that each volume remains sufficiently large?
With RELMAG = -100 (for the
With RELMAG = -100 (for the final CARD 6 that terminates the list of data sets) it will do a refinement of the parameters, but not a reconstruction. For iterations, you need to run a script that runs Frealign several times, each time doing a parameter refinement (parallelized by running several Frealign jobs with RELMAG = -100.0) and reconstruction (single run of Frealign with MODE = 0 and RELMAG = 0.0).
To speed up reconstruction, you should set FSTAT = F and also make sure there is no 3D structure read in on input. If there is a 3D structure provided on input, it is overwritten by the new 3D structure. However, the old structure on input will still be read in and used internally to calculate additional statistics, slowing down the reconstruction procedure. Normally, these additional statistics are not needed and, therefore, one should delete the 3D structure if one exists) that will then be generated on output.
3D reconstruction can also be accelerated by using the OpenMPI version of Frealign. On an 8 core system, the reconstruction should be about 3 times faster.
For 18,000 350 x 350 pixel asymmetric particles, a reconstruction should be done in less than an hour if statistics are switched off (FSTAT = F), no 3D structure is provided on input and you are running on a single core CPU. On 8 cores, this would go down to 20 minutes or less. for 100,000 particles, this would go up to a bit less than 2 hours. If you have symmetry in your particle, reconstruction times go up because each image is inserted into the reconstruction N times where N is the symmetry redundancy.