A simple infrastructure and graphical user interface (GUI) for distributed NONMEM analysis on standard network environments
Ron J Keizer(1), Anthe S Zandvliet(2), Alwin DR Huitema(1)
(1) the Netherlands Cancer Inistitute / Slotervaart Hospital, (2) NV Organon / Shering-Plough, Oss the Netherlands
Introduction: To increase processing power for CPU intensive computing encountered in PK-PD modeling, several solutions for distributive computing are available, all of which require investments in hardware, software and/or personnel.[1,2] We aimed to develop an infrastructure utilizing spare CPU cycles of desktop PC's in a standard network environment, to be set up and maintained without the need for vast hardware/software knowledge. Alongside, a graphical user interface (GUI) was constructed to use the infrastructure and to assist in performing NONMEM analyses.
Methods: The infrastructure requires a shared network-drive accessible by all clients (standard desktop PCs in a network environment), and a cron service (e.g. WinCron, free software) installed on every client. NONMEM runs are compiled locally, after which the executable file is transferred to the shared network drive and a client is assigned (manually or automatically). Runs are executed under low priority, to maintain continuous processing power for the client’s owner. Run results and additional files (MSF, tables) are transferred to a shared drive or back to the original directory. A GUI (Piraña) was developed to enable cluster distribution and to facilitate NONMEM analysis on stand-alone computers. Piraña was written in Perl/Tk and uses the PsN-core library[3] and R to generate reports of run results and basic GOF-plots.
Results: The developed infrastructure has been used successfully now for over one year by up to five modelers simultaneously. The infrastructure consisted of approximately 25 clients, and supported distribution of multiple runs to PC’s with multi-core CPUs. The system was stable and time-saving in model-development and facilitated execution of CPU intensive tasks, such as bootstrapping.
Conclusions: A NONMEM cluster infrastructure was built using a standard network environment, making use of spare CPU capacity of network-clients (standard desktop PC's). The setup could be of particular interest for small modeling groups situated in hospital or academic settings. A GUI was developed to use the infrastructure while also providing some modeling and analysis tools for NONMEM.
References:
[1] Pihlgren P. et al., PAGE 15 (2006) Abstr 1032 [www.page-meeting.org/?abstract=1032]
[2] Lindbom L., PAGE 12 (2003) Abstr 438 [www.page-meeting.org/?abstract=438]
[3] Lindbom L. et al, Comput Methods Programs Biomed. 2004 Aug;75(2):85-94
