Multicore programming for measurement: A tutorial

Manufacturers insist that the microprocessors of the future will have exponentially increasing numbers of cores for as long as Moore's Law lasts. Those of us who write programs to make and analyze measurements, like all who create soft ware as part of their work, will have to write parallel code in order to reap any benefit from new generations of processors. In the past, the primary way to program in parallel on shared memory computers was to work with threads directly. Doing so is time consuming, error prone, and best left to experts in thread programming. Over the last three years or so, several toolkits or libraries have appeared that provide generic parallel algorithms. These algorithms are implemented with threads but hide the details of threading. The programmer can use these generic parallel algorithms to parallelize code without getting involved with the pitfalls of threading. Using such toolkits or libraries is currently the best way to reap im proved measurement analysis speed on multicore processors and to future-proof that code for the upcoming generations of microprocessors with doubling and re-doubling numbers of cores.

The Programmer's View of Multicore

This tutorial focuses on multicore processors that operate like a set of traditional microprocessors, such as x86s and ARMs, that are bundled together so as to share a common main memory (Fig. 1). Other multicore architectures such as current Graphics Processing Units (GPUs) are beyond the scope of this article. GPUs essentially comprise hundreds of specialized floating point cores and present themselves to software as tens of thousands of logical cores. The resulting massive floating point capability potentially makes GPUs powerful accelerators of measurement processing and requires programming techniques that are specific to them [1]. Fig. 1.

A programmer's view of a modern multicore microprocessor.