On Fri, Aug 26, 2011 at 2:09 PM, Francis E Reyes wrote:

Isn't it parallelized in the nightlies?

Yes, Ralf and I added this a couple of months ago. There are a few more improvements that we'd like to make, such as setting 'nproc' automatically (and adjusting to compensate for system load), and fixing a couple of inefficiencies (e.g. after running a grid search to identify optimal weight, it then refines *again* with that weight, instead of saving the coordinates from the grid search result), but it is already a significant improvement.

- phenix.refine:

 - grid searches parallelized: use 'nproc' parameter to set number of CPUs    to use    - this affects bulk solvent mask optimization, and XYZ and ADP weight      optimization (in addition to TLS identification)    - can reduce runtime by 75% or more depending on settings (see manual for      details)

FYI, in practice the largest number of CPUs that the weight optimization routines will use appears to be 18 cores. (The automatic TLS identification will use as many as possible, but this won't have much additional effect.) -Nat

On 08/26/2011 09:14 PM, Nathaniel Echols wrote:

On Fri, Aug 26, 2011 at 2:09 PM, Francis E Reyes wrote:

...

Isn't it parallelized in the nightlies?

Yes, Ralf and I added this a couple of months ago. There are a few more improvements that we'd like to make, such as setting 'nproc' automatically

Maybe this Python code can be helpful: def get_nb_procs(): res = None try: # POSIX res = os.sysconf('SC_NPROCESSORS_ONLN') except: try: # {Free|Net|Open}BSD and MacOS X res = int(commands.getoutput("sysctl -n hw.ncpu")) except: res = 0 return res It is supposed to work well on Linux plus all the explicitly listed platforms. Best regards, F.

(and adjusting to compensate for system load), and fixing a couple of inefficiencies (e.g. after running a grid search to identify optimal weight, it then refines *again* with that weight, instead of saving the coordinates from the grid search result), but it is already a significant improvement.

...

- phenix.refine:

- grid searches parallelized: use 'nproc' parameter to set number of CPUs to use - this affects bulk solvent mask optimization, and XYZ and ADP weight optimization (in addition to TLS identification) - can reduce runtime by 75% or more depending on settings (see manual for details)

FYI, in practice the largest number of CPUs that the weight optimization routines will use appears to be 18 cores. (The automatic TLS identification will use as many as possible, but this won't have much additional effect.)

-Nat _______________________________________________ phenixbb mailing list [email protected] http://phenix-online.org/mailman/listinfo/phenixbb