ABSTRACT
Mole-8.5 was established to provide a customized hardware that can take full advantage of the CPU-GPU hybrid architecture to implement the multi-scale computing mode based on energy minimization multi-scale (EMMS) and discrete simulations. The communication between the nodes is important for parallel computation especially for computing nodes with multiple GPUs like Mole-8.5. Institute of Process Engineering (IPE) has tested the communication bandwidth and latency between two random nodes of Mole-8.5 using the OSU_Benchmark with a series of packages of different sizes. The Mole-8.5 supercomputer consists of 362 computing nodes. The typical nodes of Mole-8.5 are based on the GPU+CPU heterogeneous structure with a high density of GPUs, developed jointly by IPE and TYAN. The main applications running on Mole-8.5 are based on discrete simulation, which covers a variety of systems and processes, such as granular flow, emulsions, polymers and proteins, foams and micro-/nano- flows, crystals and reaction-diffusion processes.
