Re: NONMEM/PsN benchmark for SGE expansion

On 3/25/2011 4:42 PM, Ivashina, Julia wrote: > * What type of analysis is the most sensitive to an increase in > number of work nodes? As far as I know, the version of nonmem that you have installed does not support parallel processing. Legend has it that this feature will show up in the next version of nonmem (a beta version is said to exist... ;-) ). > * What should be the expected gain from increasing the number in > -threads 50 times? As you've experienced this depends on a number of factors; your distributed computing facility will bring you the largest benefits with very CPU intensive jobs. This could be models with very large datasets or PKPD models relying on numerical integration (ADVAN6, ADVAN8, ADVAN13). Smaller runs will spend a comparatively higher proportion of the runtime compiling nonmem and pushing data across the network. The way you've set up the grid file system plays a role here too. I'm not too surprised that you don't see much benefit when runtimes are as short as 5-10 seconds. > * What parts of NONMEM/PsN are the most optimized for parallel > execution? Bootstrap and VPC/NPC are the scripts that are most suited for running on a grid. Use the command line parameters -samples and -threads with the boostrap command, and -samples and -n_simulation_models (and possibly -threads) with VPC/NPC. See: http://psn.sourceforge.net/pdfdocs/npc_vpc_userguide.pdf and http://psn.sourceforge.net/pdfdocs/bootstrap_userguide.pdf > * What are the scenarios where gain from parallelization is the biggest? In my experience, your gains will be biggest when executing many long-running jobs (eg a complete model history, a bootstrap, or a NPC/VPC with the -n_simulation_models parameters) Kind regards, -- Paul Matthias Diderichsen, PhD Quantitative Solutions B.V. +31 624 330 706