Too high energy consumption is widely recognized to be a critical problem in large-scale parallel computing systems.The LogP-based energy-saving model and the frequency scaling method were proposed to reduce energy consumption analytically and systematically for other two representative barrier algorithms:tournament barrier and central counter barrier.Furthermore,energy optimization methods of these two barrier algorithms were implemented on parallel computing platform.The experimental results validate the effectiveness of the energy optimization methods.67.12% and 70.95% energy savings are obtained respectively for tournament barrier and central counter barrier on platforms with 2048 processes with 1.55%?8.80% performance loss.Furthermore,LogP-based energy-saving analytical model for these two barrier algorithms is highly accurate as the predicted energy savings are within 9.67% of the results obtained by simulation.
Heterogeneous systems with both Central Processing Units (CPUs) and Graphics Processing Units (GPUs) are frequently used to accelerate short-ranged Molecular Dynamics (MD) simulations. The most time-consuming task in short-ranged MD simulations is the computation of particle-to-particle interactions. Beyond a certain distance, these interactions decrease to zero. To minimize the operations to investigate distance, previous works have tiled interactions by employing the spatial attribute, which increases the memory access and GPU computations, hence decreasing performance. Other studies ignore the spatial attribute and construct an all-versus-all interaction matrix, which has poor scalability. This paper presents an improved algorithm. The algorithm first bins particles into voxels according to the spatial attributes, and then tiles the all-versus-all matrix into voxel-versus-voxel sub-matrixes. Only the sub-matrixes between neighboring voxels are computed on the GPU. Therefore, the algorithm reduces the distance examine operations and limits additional memory access and GPU computations. This paper also adopts a multi-level programming model to implement the algorithm on multi-nodes of Tianhe-lA. By employing (1) a patch design to exploit parallelism across the simulation domain, (2) a communication overlapping method to overlap the communications between CPUs and GPUs, and (3) a dynamic workload balancing method to adjust the workloads among compute nodes, the implementation achieves a speedup of 4.16x on one NVIDIA Tesla M2050 GPU compared to a 2.93 GHz six-core Intel Xeon X5670 CPU. In addition, it runs 2.41x faster on 256 compute nodes of Tianhe-lA (with two CPUs and one GPU inside a node) than on 256 GPU-excluded nodes.
Particle-in-cell (PIC) method has got much benefits from GPU-accelerated heterogeneous systems.However,the performance of PIC is constrained by the interpolation operations in the weighting process on GPU (graphic processing unit).Aiming at this problem,a fast weighting method for PIC simulation on GPU-accelerated systems was proposed to avoid the atomic memory operations during the weighting process.The method was implemented by taking advantage of GPU's thread synchronization mechanism and dividing the problem space properly.Moreover,software managed shared memory on the GPU was employed to buffer the intermediate data.The experimental results show that the method achieves speedups up to 3.5 times compared to previous works,and runs 20.08 times faster on one NVIDIA Tesla M2090 GPU compared to a single core of Intel Xeon X5670 CPU.