LINPACK

LINPACK
Name	LINPACK
Developed by	Jack Dongarra, Jim Bunch, Cleve Moler, Gene H. Golub
Initial release	1979
Latest release	ongoing
Language	Fortran, C, BLAS
Operating system	UNIX, IBM AIX, Microsoft Windows, Cray-1
License	permissive academic

LINPACK

LINPACK is a numerical linear algebra software library and a family of routines originally developed to solve systems of linear equations and linear least-squares problems on early high-performance computers. Conceived in the late 1970s at a research nexus involving University of Tennessee, Stanford University, and Argonne National Laboratory, LINPACK became a foundational tool for computational scientists working with matrix factorization, direct solvers, and performance measurement on architectures from the Cray-1 to modern supercomputers.

History

LINPACK originated from collaborative research by scientists at University of Tennessee, Sandia National Laboratories, Argonne National Laboratory, and Stanford University during an era marked by projects such as Project MAC and institutions like Los Alamos National Laboratory that demanded reliable numerical software. Key contributors included Cleve Moler, who later co-founded MathWorks and influenced MATLAB, and Jack Dongarra, who later contributed to BLAS and the Top500 project. Early releases targeted machines such as the Cray-1 and various CDC series systems, and integration with vendor-specific math libraries at organizations like IBM and Intel fostered broader adoption. LINPACK’s development paralleled contemporaneous efforts like EISPACK and informed standards eventually used by organizations including IEEE and ACM.

Algorithms and Components

LINPACK implements dense matrix algorithms centered on Gaussian elimination with partial pivoting, LU decomposition, and QR factorization, relying on numerical linear algebra theory advanced by figures such as Alan Turing and John von Neumann. Core components include routines for solving Ax = b via LU decomposition, least-squares via orthogonal factorization, and condition number estimation, drawing on foundational work from Gene H. Golub and William Kahan. LINPACK’s modular design separates algorithmic kernels from low-level matrix operations, enabling replacement by optimized implementations like BLAS Level 1–3 kernels. Numerical stability techniques in LINPACK reference pivoting strategies examined in research at institutions such as Princeton University and Massachusetts Institute of Technology, and leverage matrix factorizations used across computational projects at Lawrence Livermore National Laboratory.

Performance and Benchmarking

LINPACK became synonymous with numerical performance when used as a benchmark to measure floating-point throughput on high-performance systems, inspiring metrics later formalized by the Top500 project. The LINPACK benchmark exercises double-precision dense linear algebra and reports performance in FLOPS, linking machine assessments at centers like Oak Ridge National Laboratory, EuroHPC initiatives, and universities such as University of California, Berkeley. Benchmark variations, including Highly Parallel LINPACK, emerged to reflect scaling on distributed-memory systems using communication libraries such as MPI developed at Argonne National Laboratory and Los Alamos National Laboratory. The benchmark’s historical role influenced procurement and public rankings at agencies like US Department of Energy and private entities like Hewlett-Packard and Cray Inc..

Implementation and Software

Original LINPACK was written in Fortran and distributed with comprehensive routine documentation; modern adaptations include C and Fortran interfaces, wrappers for Python libraries such as NumPy and SciPy, and bindings in environments like MATLAB and R. Optimized implementations deploy platform-specific libraries from vendors such as Intel (MKL), AMD (ACML), and NVIDIA (cuBLAS) that accelerate BLAS kernels underpinning LINPACK routines. Development communities around projects like Netlib and version control at organizations akin to GitHub facilitated portability across operating systems including UNIX, Linux, Microsoft Windows, and vendor-specific systems at IBM and Cray Research.

Applications and Impact

LINPACK routines have been embedded in scientific workflows across disciplines represented by institutions like NASA, NOAA, CERN, and European Space Agency for simulations in climate modeling, computational fluid dynamics, structural engineering, and signal processing. The library’s algorithms informed numerical curricula at universities such as Harvard University and Stanford University and influenced commercial products from companies like MathWorks and IBM. The LINPACK benchmark shaped market perceptions and funding priorities at national laboratories including Lawrence Berkeley National Laboratory and influenced high-performance computing procurement policies at ministries and agencies in nations represented by European Commission research programs.

Limitations and Criticism

Critics from communities at Lawrence Livermore National Laboratory and academia noted that LINPACK-focused benchmarking can misrepresent real-world workload performance, particularly for sparse matrix problems addressed by libraries such as PETSc and Trilinos. Concerns raised by researchers at Cornell University and UC San Diego emphasize that dense linear algebra benchmarks do not capture memory-bound, I/O-bound, or mixed-precision characteristics found in applications on systems from Google and Amazon Web Services. The reliance on vendor-optimized BLAS has led to debates among stakeholders at IEEE and ACM about portability versus peak performance, prompting proposals for alternative benchmarks and suites developed at consortia including SPEC and the HPCG project.

Category:Numerical linear algebra