ICAM – I2CAM Institute for Complex Adaptive Matter

ICAM GPU Server

Need help on using the ICAM GPU Server? Check out our wiki!

Background

In Winter 2009, plans were made to begin construction of an computing cluster primarily powered by Graphics Processing Units (GPUs). The primary motivation was to take advantage of an emerging technology which may allow computational speed increases of up to 20x or more compared to CPUs for molecular dynamics algorithms. The cluster was completed in June 2010, and consists of:

1 head node
7 computing (or “data”) nodes
A 16-port gigabit switch

Hardware Specifications

Head Node (1)

The head node is made up of a 2.66GHz Intel Quad-Core processor, a Geforce 8400 GS GPU, 6 GB of memory, and three 1.5 TB SATA hard drives. Its primary tasks are to distribute computing tasks to the compute nodes and store user data.

Compute Node (7)

Each compute node is made up of a 2.66GHz Intel Quad-Core processor, 2 Geforce GTX 285 GPUs, 6 GB of memory, an ASUS P6TSE LGA motherboard, a 640 GB SATA hard drive, and an 850 Watt power supply, and takes 4U of rack space. Because each node contains 2 GPUs and 8 logical CPUs, there can be up to 14 GPU-intensive processes running simultaneously with 56 CPU-intensive processes, in full parallel.

Software Specifications

Programming Interfaces

Users will initially have access to several general and GPU-specific development packages: the GNU Compiler Collection 4.1.2, OpenMPI 1.3.3, Python 2.4, NVIDIA CUDA Toolkit 3.0 and GPU Computing SDK, PyCUDA 0.94rc, and OpenMM 1.1.1 with PyOpenMM.