Increasing Molecular Dynamics Simulations Throughput by Virtualizing Remote GPUs with rCUDA

The full-understanding of the dynamics of molecular systems at the atomic scale is of great relevance in the fields of chemistry physics, materials science and drug discovery just to name a few. Molecular dynamics (MD) is a widely used computer tool for simulating the dynamical behavior of molecules by solving Newton's equations of motion. However, the computational horsepower required by MD simulations is too high to obtain conclusive results in real world scenarios. This is mainly motivated by two factors: (1) the long execution time required by each MD simulation (usually in the nanosecond and microsecond scale, and beyond) and (2) the high number of simulations required in drug discovery to study the interactions between a large library of compounds and a given protein target. To deal with the former, Graphics Processing Units (GPUs) have come up into the scene. The latter has been traditionally approached by launching large amounts of simulations in large computational clusters that may contain GPUs on each node. However, the GPUs are targeted as a single node that only runs one MD instance at a time, which is translated into low GPU occupancy ratios. In this paper, we design a strategy to increase the overall throughput of MD simulations by increasing the GPU occupancy through virtualized GPUs. We use the rCUDA middleware as a tool to decouple GPUs from CPUs, and thus enabling time-sharing. Our results show that the use of rCUDA obtains a 2x speed-up factor compared to the CUDA counterpart version while requiring a similar power budget.