We are pleased to announce the long awaited version 2.0 of HPL. It
contains major bug fixes and accuracy enhancements that were reported by the users. The major focus of this release was to
improve accuracy of the reported results and ensure scalability of the
code on the largest supercomputer installations with hundreds of
thousands of computational cores.
We are pleased to announce the long awaited version 2.0 of HPL. It's been nearly 4 years since the last stable release and the code has served the community well. The most prominent result were the numerous entries in the TOP500 and Linpack Benchmark Report (http://www.netlib.org/benchmark/performance.pdf). We've received a number of suggestions since 2004. They resulted in major changes and accuracy enhancements. The major focus of this release was to improve the code but also keep the continuity of new runs with the existing results. The new release does not produce results that are bit-wise identical with the old results due to complete overhaul of the random number generator. The numerical properties of the results as well as runtime behavior is completely compatible with the historical data with the added bonus of having the new generator scale on the largest supercomputer installations with hundreds of thousands of computational cores. Another set of issues reported by the users was related to parallel machines with large main memory. These issues were also addressed by the fixes that went into this release. Finally, an important change was made to the way that accuracy is reported. The formula that verifies the correctness of the run has been simplified and brought in line with state of the art for numerical analysis of Gaussian elimination with partial pivoting. The code of the new release is available at http://www.netlib.org/benchmark/hpl/ under the original BSD license.