Researchers who consistently need high-throughput/high-performance cluster computing can take part in the Duke Compute Cluster by purchasing nodes. An “allocation” program for use of the cluster provides access for researchers with less persistent cluster computing requirements. To get involved, either by purchasing nodes or to be kept informed about the allocation program, contact
The Duke Compute Cluster comprises 5664 CPU cores on 457 machines that average about 32 gigabytes of RAM, with many of the nodes offering 256 and even 512 gigabytes of RAM. Because of the history and the incremental development of the cluster, the collection of nodes is heterogenous though all are in the most current Dell blade/chassis form factor (M1000 enclosure and M6xx blade). The cluster uses the SLURM scheduler. The Duke Compute Cluster consists of machines that the University has provided for community use and that researchers have purchased to conduct their research. The University houses and supports all the machines for their useful life. There are times when members require more exclusive access to their own node, and groups who have bought nodes for the cluster have “high priority” access to them when they require it. Jobs submitted with high priority run only on the nodes that members have bought. Of course, most members of the Duke Compute Cluster usually do not run jobs 100% of the time, and thus the communal nature of the low priority scheme, by which members make resources available to others when they are not requiring it, increases the efficiency of the overall system, and often provides researchers much more compute when they need it. Since April 2015, Duke Research Computing has offered another means of becoming involved with the Duke Compute Cluster through an “allocation program,” which grants cluster access for a limited duration in support of computationally intensive projects. These opportunities support researchers who otherwise may not have funding for more substantial investment for high priority access or whose projects are more discretely bounded.