Floating-point applications from Hollywood to Wall Street have long faced the challenge of providing both great performance and exactly the same results from run to run, or in other words, reproducible results. While the main factor causing a lack of reproducible results is the non-associativity of most floating point operations, there are other contributing factors such as runtime, selectable optimized code paths, non-deterministic threading and parallelism, array alignment, and even the underlying hardware floating-point control settings.
In this article for Intel® software tool users and programmers, we outline how to use the Intel® Math Kernel Library (Intel® MKL) and Intel® compiler features to balance performance with the reproducible results applications require. These new reproducibility controls in Intel® Parallel Studio XE 2013 help make consistent results from run to run possible.
Read the full article from Intel’s Parallel Universe magazine here.
