High-performance computer cluster - Landau
We built a cluster of computers, called Landau, which is up and running in the IT department at the University of Massachusetts at Lowell (University Data Center). Landau has 12 CPU computational nodes and one master node, each node equipped with two Intel Xeon E5440 (2.83GHz) quad core CPUs. This is 8 computational cores per node, 104 computational cores in total. Each node has 4 x 1GB of memory. Nodes are interconnected via 1 Gigabit Ethernet network. The cluster is running Fedora 7 Linux. Landau is primarily used for molecular simulations using standard
packages for all-atom Molecular Dynamics (MD) simulations, such as NAMD, Gromacs and CHARMm. The roughly estimated raw computational power of Landau is ~1.2 TFlops.
GPU-based computational nodes
4U rack-mounted computational nodes located in the same server rack as Landau cluster. Each node is equipped with four NVidia GeForce GTX 480 GPUs (total of 8 GPUs or 8x480 ALUs in two nodes), Intel Core i7 930 Bloomfield quad-core CPU (2.8 GHz) and 6 GB of DDR3 1600 MHz memory. The raw computational power of one GeForce GTX 480 is ~1.35 TFlops (~675 GFlops) for single (double) precision computations. This corresponds to the total computational power of one node of ~5.4 TFlops (~2.7 TFlops). To assemble the GPU nodes, we used Supermicro SC747TQ-R1400B chassis to provide sufficient power and airflow for 4 x 250 Watt GPUs and ASUS P6T7 WS Supercomputer motherboard with four PCI Express 2.0 x16 slots. These CUDA-enabled GPU empowered computational devices are running Ubuntu 10.04 Linux. The GPU nodes are used mostly to run biomolecular simulations which
employ coarse-grained descriptions of biomolecules (SOP-GPU package) and for all-atom simulations in implicit solvent (CHARMm software).
In addition, to Landau, we also use high-performance workstations in our computer laboratory equipped with GPUs of different generations (G200 and Fermi). Among these are 2 high-performance GPU workstations, each with 3 NVidia GeForce GTX 295 graphics cards and one NVidia Quadro NVS 295 card for display support. Since each GeForce GTX 295 has 2 GPUs on board, the total number of CUDA-enabled graphics processors in each workstation is 6, and the total number of ALUs is 6x240. These workstations are based on ASRock X58 Supercomputer motherboard, equipped with Intel Core i7 920 Nehalem CPU (2.66GHz); they have 6GB of 1333MHz DDR3 memory and supported by 1600W X3 power supply. The peak computational power from one workstation is ~5.4 TFlops (single precision arithmetic). Workstations are running Ubuntu 9.10 Linux. These computing devices are used for Langevin simulations of proteins (SOP-GPU package), and for new software development and testing.
The total raw computational power of the GPU-based solutions in our research group exceeds 20 TFlops.