differences between command line and GUI?
Dear PHENIX developers, I ran a job through the GUI, and then what I thought was the same job on our cluster. The job from the cluster has Rfree that is 0.7% higher. I'm totally at a loss to figure out what is causing this. It is the same version of PHENIX (1.7-650). I then took the .eff file from the GUI job and used that as the params file for the command line job but still get the big difference in R free. Any suggestions? Best regards, Kendall Nettles
Hi Kendall, I am guessing that this is an autobuild job, and that the machines used in the cluster are different than the machines used from the GUI. If this is the case, then I am not too surprised, as autobuild makes many decisions that are either/or (place a chain here or not) rather than continuous (move this atom here) so that tiny numerical differences can lead to significant differences in the final model. Now if you run from the GUI or command line and all machines are the same...then it should be the same, because the GUI is creating a parameters file and then running that command file. Autobuild also uses a random seed to generate trial positions for new chains, so if you re-run your job with i_ran_seed=xxxx, varying xxxx, you will get different results as well. I hope that helps! -Tom T
Dear PHENIX developers,
I ran a job through the GUI, and then what I thought was the same job on our cluster. The job from the cluster has Rfree that is 0.7% higher. I'm totally at a loss to figure out what is causing this. It is the same version of PHENIX (1.7-650). I then took the .eff file from the GUI job and used that as the params file for the command line job but still get the big difference in R free. Any suggestions? Best regards,
Kendall Nettles
_______________________________________________ phenixbb mailing list [email protected] http://phenix-online.org/mailman/listinfo/phenixbb
On Mon, Mar 28, 2011 at 5:59 AM, Kendall Nettles
I ran a job through the GUI, and then what I thought was the same job on our cluster. The job from the cluster has Rfree that is 0.7% higher. I'm totally at a loss to figure out what is causing this. It is the same version of PHENIX (1.7-650). I then took the .eff file from the GUI job and used that as the params file for the command line job but still get the big difference in R free. Any suggestions?
Which program? Try running 'diff' on the .eff files for jobs that produced different R-frees and see what the exact differences are. If you were running phenix.refine, there are two points where the parameters diverge in 1.7: 1) The command-line version reads in an additional set of parameters for real-space refinement, modifying the defaults (since these are used elsewhere for very different purposes, we need two sets of defaults). The GUI doesn't do this, because I wasn't aware of it at the time. Next version will use the same parameters for both (as soon as I fix it...). 2) The GUI sets the maximum RMS(bonds) and RMS(angles) for weight optimization much lower than the defaults for the command-line version, which I felt were unreasonably high (and led to geometry problems for some structures). It appears that the defaults are now closer to what I set, so it may be time to remove this discrepancy too. Either of these could explain the difference in R-free. However, as Tom points out, many programs will come up with a slightly different answer depending on where they are run, often for reasons that don't make sense. If you switch between Mac and Linux, you'll almost certainly see this happen regularly. -Nat
participants (3)
-
Kendall Nettles -
Nathaniel Echols -
Thomas C. Terwilliger