A Report on molecular dynamics simulation machine: Anton

IntroductionThe process of drug development in the field of medical science needs detailed understanding of the interaction between the protein molecules and the drug being developed. This essentially provides the need for development of a computer based simulation system for the study of the molecular dynamics to understand by visualizing the motion of these molecular systems precisely in an atomic level detail. The molecular dynamics (MD) simulation of these bio molecular systems would require calculations of the forces exerted on tens and thousand of atoms in few microseconds which is beyond the computational capability of the currently available super computer technology. A massive parallel machine called Anton is being developed to accelerate the molecular level simulation process by several orders. Anton which is a specialized computation engine named Anton van Leeuwenhoek (father of microscopy) has been implemented as a single ASIC comprising of 512 processing nodes. It formulates a three dimensional torus by spatially decomposing and distributing the molecular system among the interconnected nodes. A novel parallel algorithm is designed to attain a high level of arithmetic density which is capable of reducing the communication time among the chips with the aid is specifically designed hardware system. Design of these hardware systems for Anton is specifically to examine the trajectories of every single molecule involved in the biological and chemical processes which may have required one to run several thousand computers simultaneously. Several other projects which were involved in development of computational capabilities were limited by the Amdahls law and other communication bottlenecks which prevented them from using several such chips to extend the simulation capabilities beyond the micro second timescale. Computation of Molecular Dynamics in AntonThe methodology of computation of molecular dynamics in Anton is done in four discrete steps which are based on the laws of classical physics. The first most steps involves dividing time discrete and uniform time steps in the order of few femtoseconds. These time steps are then subdivided into two phases to perform distinctive tasks of calculating force on each particle in the system and Integrate it to predict the position and velocity of the particle. Force calculation is done based on available models to predict the three force components; bond forces, van der waals forces and electrostatic forces as a function of spatial coordinates of all atoms. The computational workload associated with the calculation of electrostatic interaction is reduced by Anton with the help of k-space Gaussian split Ewald method by dividing it into two sub components range limited interaction and long range limited interactions. A specially designed algorithm called the NT method is used to parellize the range limited interaction by achieving both asymptotic and partical reduction in required inter processor communication bandwidth through assigning spatial assignment of particles to nodes, but that often compute the interaction between two particles using a node on which neither particle resides.Integration of the calculated forces is then executed to predict the atomic position and velocity from the sets of differential equations that describes the motion of the atoms. Integration of these equations to predict the molecular dynamics requires sophisticated algorithm that could handle adjustments in global temperatures and pressure in a more accurate manner. With the aid of constraints during the integration process the simulation can be further accelerated.Requirement of Specialized HardwareAnton requires a simulation speed of 1000 times faster to simulate the interactions of few milliseconds in a couple of months which is possible by completing the time steps within every few microseconds. The sequential dependency in an MD simulation makes it essential for development of specialized hardware that can accelerate the individual time steps with the aid of architectural enhancements that would reduce both computational and communicational latency. This is balanced by limiting the memory requirements in Anton to almost around 3.2 KB per node by utilizing only SRAMs and small L1 caches. The calculation of electrostatic and van der waal forces accounts to almost 90% of the computational time in a general purpose supercomputing hardware which contradicts to the Amdahls law. Hence a specialized hardware has to be essentially developed to provide significant silicon area for accelerating the other computation tasks of calculations of force fields and its integration.Architecture of the systemNodes are the building blocks of the system architecture, each one of which comprises of an MD specific ASIC, DRAM and 6 different ports for the system wide interconnection network. A three dimensional torous topology is used to connect logically identical nodes. ASIC which are clocked at 400 MHz frequency further consists of 4 major subsystems. 1. High throughput Interaction subsystem (H-TIS):This subsystem performs the task of charge spreading and force interaction by applying a massive parallelism. It does the bulk calculation or the molecular dynamics by its capability of computing the combined electrostatic and van der waals interaction between a pair of atoms at every cycle with help of an array of 32 pairwise point interaction modules (PPIMs). 2. Flexible subsystem: This controls the ASIC and is also responsible for carrying out the calculation of bond forces, Fats Fourier Transform and integration. It is divided into four slices each of which comprising remote access units responsible for data transfer and computation. It also consist of a correction pipline, interconnection for flexible subsystem components and an interfacing unit to transferring data to the communication subsystem. 3. Communication subsystem: It provides a low latency high speed communication pathway for ASIC and their subsystems. The speed goes upto 5.3 GB/s between the chip torus. This subsystem provides a support for the efficient multicast, flow control, rate metering and an access to the external host computer to provide I/O for the simulated data.4. Memory Subsystem:These subsystems provide an access to the DRAM and also support the synchronizaition and accumulation in addition to its read and write capabilities. The operations helps in reducing the load over its other subsystems and provides a capability to the Anton to simulate chemical systems of billions of atoms.Performance and Accuracy:Anton can achieve a high simulation rate of scale of 14,500 nanoseconds per day which can simulate the interaction of milliseconds within two months. It can achieve a relative rms force error of 1.5 x 10-4 which is the closest to 10-3 of the other MD computational hardwares.