Guidelines for Writing System Software and Tools Requirements

Terms and Definitions

Terminology

• Platform: Generic term for a parallel or clustered computing system. Unless explicitly noted otherwise, capabilities apply equally to shared-memory, distributed-memory, and SMP systems. Note that a particular vendor may implement multiple platforms, so an element described as platform-specific may be supported by a vendor in more than one way.

• PE (Processing Element): Generic reference to the basic hardware unit on a given platform; may correspond to "node", "processor", "CPU", or "PE", depending on the system.

• API (Application Programming Interface): Syntax and semantics for invoking services from within an executing application. All APIs must be available to both Fortran and C programs, although implementation issues (such as whether the Fortran routines are simply wrappers for calling C routines) are up to the supplier.

• Standard (as applied to APIs): Where an API is required to be consistent across platforms, the reference standard is named as part of the capability. The implementation must include all routines defined by that standard (even if some simply result in no-ops on a given platform).

• Published (as applied to APIs): Where an API is not required to be consistent across platforms, the capability lists it as "published," referring to the fact that it must be documented and supported, although it may be vendor- or even platform-specific.

• Current standard: Term applied when an API is not "frozen" on a particular version of a standard, but should be upgraded automatically by vendors as new specifications are released (e.g., "MPI version 1.1" refers to the standard in effect at the time of writing this document, while "current version of MPI" refers to further versions that take effect during the two-year lifetime of this document.

• Fully-supported (as applied to system software and tools): A product-quality implementation, documented and maintained by the HPC machine supplier or an affiliated software supplier.

• XXX-compatible (as applied to system software and tool definitions): Requires that a capability be compatible, at the interface level, with the referenced standard, although the lower-level implementation details may differ substantially (e.g., "DCE/DFS-compatible" means that the distributed file system must be capable of handling standard DFS requests, but need not conform to DFS implementation specifics).

• Single-point control (as applied to tool interfaces): Refers to the ability to control or acquire information on all processes/PEs using a single command or operation.

Prioritization Levels

• Baseline Development Environment (BDE): Capabilities that are needed by an overwhelming majority of HPC user sites. The recommendation of the task force is that the entire BDE be included on all procurements for parallel and clustered computing systems.

• Desirable Level 1 (D1): Capabilities that are important to a significant number of HPC user sites, but not to an overwhelming majority.

• Desirable Level 2 (D3): Capabilities that are important to a significant number of HPC user sites, but involve significant extra implementation effort or embody a better way of doing something that can already be accomplished (albeit crudely) using the BDE or D1 elements.

• Desirable Level 3 (D3): Capabilities that are important to many HPC user sites, but can be interpreted as "value-added" rather than essential needs.

• Omitted: A number of other capabilities were considered by the task force. These were omitted from the final document either because the technology is not sufficiently advanced to permit a clear specification of the requirement, or because the group felt that the requirement did not apply to a large enough cross-section of the HPC user community. They provide important indicators of future needs, however. See the supporting documentation for details.