Linux Server Appliance EDI Linux Software: Message passing is an important tool for sites relying on HPC on Linux-based systems. It enables large data sets to be tackled with ease and often represents an important adjunct to threading-based solution design.

---

by Andrew Binstock, principal analyst, Pacific Data Works LLC. Intel Corp.

Message passing is an important tool for sites relying on high-performance computing (HPC) on Linux systems. It enables large data sets to be tackled with ease and often represents an important adjunct or alternative to thread-based solution design. Integrating message passing and threading has its challenges, however.

High-performance computing is characterized by running multiple tasks in parallel. These tasks tend to be similar in nature and can be run on different systems (as in a cluster or a grid), different processors (on a single system) or even different execution pipelines (as on processors with Hyper-Threading Technology). HPC systems add computing power by the addition of discrete computing engines (be they blades, nodes, or processors) rather than by upgrading existing processors. This aspect underscores one of the salient aspects of HPC: the number of executing processes can vary, and so software must be architected to scale dynamically across the number of available execution units.

One way to write software that dynamically takes advantage of a variable number of execution resources is to use OpenMP, a set of threading interfaces and tools that greatly simplifies threaded programming. Using OpenMP pragmas (in C/C++, or directives in Fortran), you can generate an executable that determines at run time the appropriate number of threads to create and then distribute a workload over these threads dynamically. OpenMP's other advantages, such as portability across a wide set of platforms, are detailed in a series of articles on this Web site. These articles also explain how to get started with OpenMP. For additional information, go to the OpenMP home page at www.openmp.org.

To read more, click link below.