For a better experience, click the Compatibility Mode icon above to turn off Compatibility Mode, which is only for viewing older websites.

Fair Share Allocations on Proteus

The Drexel Cluster consists of a login node and compute nodes, interconnected by both standard and fast networks. The rationale in offering this resource is to free researchers from the need to purchase entire clusters and instead focus their equipment resources on compute hardware and to avail them of the systems administration services the Cluster requires.

The base unit of compute hardware is termed a "share". We currently offer for purchase both Intel- and AMD-based shares. Current pricing is $25,500 for an E5-2670-based 64-core Intel share, which consists of four 2-socket nodes with 64 GB RAM each, and $18,000 for an Opteron-6378 128-core AMD share, which consists of two 4-socket nodes each with 256 GB RAM. Both share types are equipped with QDR Infiniband. These prices reflect the bare cost of the node and a uniform allocation of the noncompute infrastructure (racks, switches, cables, etc.) and are reevaluated by the Board at 6-month intervals. Our current HPC hardware vendor is Dell.

In order to become a contributor, in FY14, a PI must contribute at least one share. Monies in the amount required based on the number and types of shares desired must be transferred into the main URCF fund to execute a purchase. Purchases are made on an ad hoc basis but are anticipated to conform to a regular quarterly or semiannual schedule at steady-state.

Hardware purchased with a share contribution is not itself for exclusive use by the contributing PI, but instead entitles the contributing PI to a share of the total cluster proportional to the fraction of nodes he/she contributed.

The lifetime of a share is at least three years, beyond which nodes are evaluated by the system administrator and expected to be decommissioned after five years. When decommissioned, the share associated with that node is released from the PI.

The number of shares determines the PI's group's entitlement to cluster resources, according to a fair share scheduling algorithm.

Please email us for details.