Yan Li, Nianhang Chen and Simon Zhou*
Celgene Corporation
Objectives: To evaluate efficiency gain of the two parallel computing methods in resolving complicated popPK and popPK/PD models.
Methods: Subjects from 8 clinical studies were included the popPK and popPK/PD models. The PK model has 3 compartments, with saturable elimination from the central compartment, saturable distribution between the central and first peripheral compartments, and linear distribution between the central and second peripheral compartments. The PK/PD model has 8 compartments, mimicking physiologic processes of neutrophil precursor cell maturation. There were 18 parameters (10 for fixed effect and 8 for random effect) in the PK model and 11 parameters (6 for fixed effect and 5 for random effect) in the PK/PD model.
Results: Use of parallel computations with 8 cores running on the same computer improved the speed significantly. In PK modeling, maximum efficiency was achieved when the numbers of nodes were set to 4 for FPI method, reducing running time by ~55% (from 21 min to 9.1 min), while further increase the nodes to 8 increased run time to 13 min. The maximum efficiency was achieved when the number of nodes was set to 8 for MPI method, reducing running time by ~70% (from 21 min to 6.1 min).
In PK/PD modeling, maximum efficiency was achieved when the numbers of nodes were set to 8 for FPI method, reducing running time by ~74% (from 266 min to 69 min), and the maximum efficiency was achieved when the number of nodes were set to 8 for MPI method, reducing running time by ~75.8% (from 266 min to 64 min).
FPI methods were also tested on multiple computers. Similar results were observed when the slower computer was configured as the master, and the faster computer was configured as the slave. Interestingly, when the faster computer was configured as the master, and the slower computer was configured as the slave, instead of increasing the computation efficiency, the overall performance of FPI parallel computation was worse than the performance on each individual computer.
Conclusions: Both FPI and MPI can significantly improve the computation efficiency. Due to the inherent difference between these two methods, the improvement of the speed is more proportional to the number of processors (cores) from MPI than that from FPI. For complicate PK/PD models, FPI and MPF provided similar efficiency. For FPI method across multiple computers, the computer with faster CPU should be always set as the master with the slower CPU as slaves.
Reference: PAGE 22 () Abstr 2869 [www.page-meeting.org/?abstract=2869]
Poster: Other Modelling Applications