Columbia High Performance Computing clusters
- Former user (Deleted)
- John S. Pellman
The Research Computing Services group maintains a number of HPC clusters for the use of the entire university. These clusters all have high speed Infiniband networking and parallel file systems which make them uniquely suitable for certain workloads within the services offered by Columbia. The complete list of the shared clusters is available here.
Scheduling in these clusters is determined using a model of node ownership, where groups purchase nodes and receive a scheduling priority based on their purchase as a percentage of the overall number of purchased nodes. Node owners are guaranteed to have access to their purchased nodes within 12 hours. The aggregate purchases made by Zuckerman are all relatively small next to those made by more computationally intensive groups so it will almost always be better to schedule jobs shorter than 12 hours as it allows access to nodes owned by other groups.
Access to the clusters
A number of labs in Zuckerman have purchased nodes in the various clusters, check with your lab manager to see if your lab owns nodes in any of these clusters. Additionally, the Institute has purchased a single node in the Habanero and Terremoto clusters, which grants all Zuckerman researchers (who have a uni) access to a general ZI group on those clusters.
To get access to the clusters send an email to the Zuckerman research computing group with the cluster you wish to join and the group you're joining. Please include a cc to your lab manager as an indication of their approval if you're joining a specific lab's share.