NC State HPC Faculty Partner Program Researchers purchase compatible HPC hardware and any specialized or discipline-specific software licenses. NC State Office of Information Technology (OIT) provides space in an appropriate and secure operating environment, all necessary infrastructure (rack, chassis, power, cooling, networking), and the system administration and server support. In return for infrastructure and services provided by OIT, when partner compute resources are not being used by the partner they are available to the general NC State HPC user community.
	Partner Program Computational Hardware Partner program hardware options are compatible with the general HPC hardware being operated by OIT. Compatible hardware allows limited systems staff to effectively support a large number of systems - since the systems effort required to manage the cluster increases very little with additional compatible hardware. Two general environments are currently available for partners: a distributed memory linux cluster environment based on IBM BladeCenter hardware and a shared memory environment based on IBM POWER5 hardware. Distributed Memory Linux Cluster Partners can choose between Intel dual/quad-core Xeon based or AMD Opteron based Linux cluster compute nodes. These compute nodes have two multicore processors (currently dual and quad core Xeon and dual core Opteron) and are typically configured with 2GB of memory per processing core and a 36GB disk. Partner cost for Linux cluster compute nodes is the actual cost of the blade with three years of maintenance included. OIT provides chassis space and all other necessary infrastructure for Gigabit Ethernet connection to the henry2 cluster for Xeon blades or to the tim cluster for Opteron blades. Myrinet and InfiniBand low latency interconnect options are available for partner compute nodes (with additional cost). Shared Memory System Partners can add additional p575 nodes to the HPC POWER5 shared memory system. These are 8-processor nodes typically configured with 32GB of memory, two 73GB SCSI disks, and fibre channel ports for connection to gpfs disk arrays. Partner cost for p575 nodes is the actual node cost with three years of maintenance. OIT provides the p575 rack space and all other necessary infrastructure to include the node within the POWER5 system environment. Management of Partner Compute Resources Both the distributed memory Linux cluster and the shared memory system use Platform LSF for resource management and scheduling. LSF fair share scheduling is used to provide equitable access to compute resources. Partners have a dedicated LSF queue that provides access to their compute resources. Also, the LSF fair share value for partners reflects their participation in the overall resource. This allows partners to utilize their resources through their exclusive queue or to utilize general resources with a higher priority (based on fraction of overall resources owned by the partner). All access to compute nodes is through LSF. Separate, shared login nodes provide access to all HPC compute resources. Advantages The HPC compute node Partner Program offers compelling advantages for both the faculty partner and for the university. Partner Advantages (services provided by university) + secure space + power (including UPS and diesel generator) + cooling + rack (including rack power distribution) + network infrastructure (including message passing network for distributed memory nodes) + system administration and maintenance + priority access to additional compute resources + access to shared storage and file systems + access to university licensed software (compilers, debuggers, optimized math libraries, performance analyzers, ...) + system and computational science support from HPC staff University Advantages + multiplies resources provided by university HPC investment + increased HPC resource utilization yields more efficient use of university-wide research computing dollars + scaling benefits reduce university-wide cost of HPC facilities (a few large power and cooling units vs. many small power and cooling units) + scaling benefits reduce university-wide cost of HPC system support (incremental system administration and maintenance load for compatible hardware is very small - that is it takes nearly the same work to operate an 8-processor cluster as it does to operate a 100-processor cluster)
	Partner Program Storage HPC Partners purchase storage which is made available from HPC resources and is operated and maintained by HPC staff. There are three storage service options available: Network Attached Storage This service uses a storage server running NFS to export storage to all HPC nodes. This type storage service is suitable for use by serial and small parallel jobs (provided the parallel jobs are not using MPI-IO). Parallel File System This service uses several storage servers (typically four) running GPFS to provide a parallel file system available to all HPC distributed memory nodes. This type storage service is suitable for use by all jobs including large parallel jobs and jobs using MPI-IO. Network Storage with Hierarchical Storage Management This service uses a storage server running NFS to export stroage to HPC login nodes. Files which have not been recently accessed are migrated to tape - while their directory information remains on disk. Any access of the file automatically recalls it from tape. This type storage service is suitable for storage of many large, infrequently used datasets. This storage service is not suitable to use as working space for running jobs. All three services include tape backup with a default policy of retaining one backup copy of each file. Partners can specify other policies (eg retain all versions for last two weeks) and resulting costs are passed through to the partner.