ACADEMIA

The Texas Advanced Computing Center (TACC) at The University of Texas at Austin announced it has released software for industry, academia and the scientific community that helps supercomputers run more efficiently. The software, named GotoBLAS, contains routines that provide standard building blocks for performing basic vector and matrix operations. These microkernel implementations are highly optimized for individual architectures and the GotoBLAS library provides extremely efficient implementations that can be used in the development of high quality scientific software. The software library speeds up a key linear algebra component frequently used during the supercomputing process and has been used to greatly increase the performance of a variety of scientific applications. It has increased the speed of some applications by as much as 50 percent. The technology is available to commercial users for a licensing fee. A portion of proceeds from commercial licenses will be reinvested in future development of the software. TACC also is offering licenses at no charge and will make the source code available to accelerate development to the academic and scientific communities. Some of the world’s fastest computers use the technology, developed by TACC research associate Kazushige Goto, to enhance their performance for benchmarking tests. He began work on the code about a decade ago as a hobby when he was a patent examiner in Japan. The process is a painstaking one which Goto writes line-by-line. “Goto is a singular talent in performance programming, and we are pleased to finally offer his valuable software via both open source licenses to the research community and commercial licenses to the technology providers that desire to include his software in their commercial products,” said Dr. Jay Boisseau, director of TACC. “This dual-license strategy supports the research community and the private sector, while ensuring that Goto’s work is also supported so that he can continue to provide the highest performance basic linear algebra subroutines (BLAS) library in the field.” Goto’s development approach is to focus on writing the critical mathematical operations in assembly language to extract the highest floating-point performance possible. This process requires intimate knowledge of the underlying processor architectures. Because Goto’s programs can more efficiently solve complex linear equations, they can offer decreased solution times to a range of computational science and engineering problems. The software is generally applicable to any scientific computing discipline which requires BLAS and the need for highly-optimized libraries is paramount for achieving high performance on modern microprocessors. “A key component in developing scientific applications that run efficiently on current terascale and future petascale systems is to utilize math libraries that deliver the highest performance possible on an increasingly complex offering of architectures,” said Dr. Karl Schulz of TACC’s High-Performance Computing (HPC) group. “As we continue to transition to multi-core, multi-socket environments as the building block for our HPC systems, Kazushige’s ability to isolate and leverage the strengths of the underlying hardware will become increasingly important in future system deployments.” "We often call Kazushige the ‘Michael Jordan’ of high-performance linear algebra kernels, to his great embarrassment,” said Robert Van De Geijn, Professor of Computer Sciences at the university. “He combines scientific insight into the interaction between architectures and algorithms with an extraordinary ability to engineer the fastest implementations. The scientific computing community is very fortunate that he makes his libraries available." The software is designed for all levels of chipset architecture including the Power 4, Power 5, Opteron, Blue Gene/L architectures, the Pentium 4/Xeon (32bit and 64bit), and Itanium 2 architectures. To download a copy of GotoBLAS for academic, research, or personal use, please visit its Web site. For additional information on how to obtain a commercial license, please contact: software@otc.utexas.edu.