We
have our own in-house 98-node Beowulf computing cluster. Each node has
a 3.06GHz Pentium IV processor with either 1.0GB or 4.0GB of memory.
The nodes communicate via gigabit ethernet switches (HP Procurve). We
have installed VASP 4.06, SIESTA 2.0, and our own molecular dynamics
codes developed with our collaborators at Argonne National Laboratory.
Parallel computations are performed using MPICH2.
The MD code works in parallel by slicing the system into smaller segments. Each segment is handled by a separate processor. Interations between atoms residing on different nodes are handled by message passing. For short-ranged interactions, this means that a node only needs to frequently communicate with two other nodes that are dedicated to neighboring slices of the system. This is extremely efficient, and we find that our running time of our code scales nearly linearly with the number of nodes.
The MD code works in parallel by slicing the system into smaller segments. Each segment is handled by a separate processor. Interations between atoms residing on different nodes are handled by message passing. For short-ranged interactions, this means that a node only needs to frequently communicate with two other nodes that are dedicated to neighboring slices of the system. This is extremely efficient, and we find that our running time of our code scales nearly linearly with the number of nodes.