Parameter Estimation for System

SHARMA, SIDDHARTH. Parameter Estimation for System Biology Models on GPU Clusters . (Under the direction of Dr. James Tuck.) In this work, we consider the parallelization of interval based parameter estimation with uncertainty propagation for non-linear ODE models. This approach is especially important in systems biology research as it addresses the uncertainty often present in data recorded from biological experiments. As systems biologists model larger systems, parameter estimation quickly becomes intractable due to the large space of possible solutions that need to be evaluated in the presence of uncertainty. Efficient parallel computing is necessary to solve these problems. The work outlined here can be broken down into three stages and outcome of each stage can be seen as a contribution. First, the interval analysis based parameter estimation workload is characterized which helps in identifying the nature of available parallelism in the workload. Next, the insights gained from the characterization step are used to design and compare two very different approaches to estimate parameters on a single GPU. GPU approaches exploit the task level parallelism in the workload by maintaining a task-queue on the GPU. Optimizations are proposed to efficiently manage and spread workload within a GPU. Together, the first two stages enabled the development of scalable multi-GPU implementation of interval based parameter estimation, which is the third contribution of this work. The key design highlight here is a novel fully dynamic load balancing technique to share pending work between GPU task-queues over an MPI communication layer. The GPU implementations achieve a speedup of 16x on a single Fermi GPU and 156x on a cluster of 16 Fermi GPUs when compared to a 8-core CPU implementation. © Copyright 2014 by Siddharth Sharma