News & Events
Navy DSRC Xeon Phi Nodes Available on iDataPlex Systems
Hybrid computing is acknowledged to be the future of High Performance Computing, and as such, the Navy DSRC is pleased to announce the availability of Intel Xeon Phi accelerated nodes on the IBM iDataPlex systems HAISE and KILRAIN. Intel Xeon Phis, also known as many integrated cores (MICs), can each support up to 240 threads of concurrent execution. A detailed Intel Xeon Phi Guide is available on the Navy DSRC website at: http://www.navydsrc.hpc.mil/docs/xeonPhiGuide.html.
Users will be able to access the nodes via the 'phi64' queue on each system. PBS jobs submitted to the 'phi64' queue will continue to be charged based on the usage of a standard compute node (# of nodes * 16 cores * # of hours). If you would like access to the phi64 queue, please send a brief note requesting access to firstname.lastname@example.org.
These nodes are comprised of both a compute node and an accelerator node. Each accelerated compute node contains two 8-core 2.6 GHz Intel Sandy Bridge processors, identical to those in the rest of the system, and 64 GB of memory. Each Xeon Phi co-processor node contains two Intel Xeon Phi 5110P co-processors, each composed of 60 cores and 8 GB of internal memory. Each Xeon Phi core supports up to 4 execution threads, allowing for potentially extensive parallelism.
In order to properly support compilation of Intel Xeon Phi codes, the Intel Compiler Suite and Intel Math Kernel Library (MKL) modules will both be defaulted to version 13.1 on Haise and Kilrain.
Please note: Intel Xeon Phi nodes on Haise and Kilrain currently only support offload mode and native mode.
In offload mode, code runs on the standard compute node portion of a Phi node and offloads segments of execution to the co-processor portion. For example, an MPI code with offload directives within it could run on multiple accelerated nodes, using the Intel Xeon Phi portion of each node to potentially speed up calculations.
In native mode, code runs directly on the co-processor node. Currently, MPI codes running in native mode are limited to a single Phi node. However, in the next several months Mellanox should release an update to its version of OFED, which will support native execution of MPI codes across multiple Intel Xeon Phi nodes.
For more information about developing code for the Intel Xeon Phi, including webinars, please see: http://software.intel.com/mic-developer
Users are invited to report problems and direct requests for unclassified assistance to the Consolidated Customer Assistance Center (CCAC) at 1-877-222-2039 or by E-mail to email@example.com.