A hybrid parallel algorithm for computer simulation of Electrocardiogram based on a CPU-GPU cluster

Abstract

Biological computations like Electrocardiological modeling and simulation usually require high-performance computing environments. This paper introduces an implementation of parallel computation for computer simulation of Electrocardiograms (ECGs). We realized the parallel computation for computer simulation of ECGs on a CPU-GPU cluster using a hybrid parallel algorithm with the parallel program development tools--MPI, OpenMP, and CUDA. Furthermore, we proposed a load-prediction static scheduling and load-prediction dynamic scheduling to achieve efficient process-level and thread-level scheduling, respectively. Compared with traditional static and dynamic scheduling, our scheduling schemes are more efficient. In our research, we achieved a speedup of 55.1 using four PCs for the computer simulation of ECGs. This study demonstrates that the cluster can provide a cheap and efficient environment for parallel computations in biological modeling and simulation studies.

Load-prediction scheduling algorithm for computer simulation of electrocardiogram in hybrid environments

“…Likewise, we would like to allocate the light workloads to the accelerating component with weak computing ability. For simplicity, we quote the research conducted previously to make a hypothesis stating that the CPU's computing ability is weaker than that of the GPU (Shen et al, 2009(Shen et al, , 2012(Shen et al, , 2013. Therefore, our approach would allocate the heavy workloads to GPU and assign the light workloads to CPU.…”

“…Thereafter, considering that the GPU is unsuited for branch structure, Shen et al proposed a GPU-based algorithm that concentrates on eliminating branches in computation and optimizes the calculation of electric potentials by means of load-prediction (Shen et al, 2012). To fully utilize all accelerating components of a computer, Shen et al employed parallel computation for computer simulation of ECGs on a CPU-GPU cluster using a hybrid parallel algorithm with parallel program development tools-MPI, OpenMP, and CUDA (Shen et al, 2013). That is the basis of the paper about the original concept of LoadPrediction Scheduling (LPS).…”

“…This is due to the following reasons. When LPS without SWM conduct the experiment on a new runtime environment, the parameters are the same as the program described in (Shen et al, 2013). When we set the position G and P as a constant, i.e., 17, in LPS's (without dynamic-scheduling window) program in our previous work (Shen et al, 2013), it implies that the scope of GPU's computation is 17 times larger than CPUs'.…”

“…When LPS without SWM conduct the experiment on a new runtime environment, the parameters are the same as the program described in (Shen et al, 2013). When we set the position G and P as a constant, i.e., 17, in LPS's (without dynamic-scheduling window) program in our previous work (Shen et al, 2013), it implies that the scope of GPU's computation is 17 times larger than CPUs'. As shown in Table 1, LPS (without dynamic-scheduling window) finished the tasks in 166-167 s while the GPU completed its tasks in 169 s. When we equipped the SWM algorithm to the LPS, the workloads are allocated in a more balanced way.…”