Libraries

The use of scientific libraries to achieve a measure of portable performance has been used across many earlier computational platforms. The ability to use higher-level abstractions via libraries allows developers to concentrate their effort on algorithmic development, freeing them from having to devote considerable effort to maximizing performance for many mathematical primitives. The most popular scientific libraries include packages designed to solve problems in linear algebra (both dense and sparse), compute fast Fourier transforms (FFT), multigrid methods, and initial value problems for ordinary differential equations, along with other examples. Some of the best known scientific libraries include:

These tasks often represent a good measure of the computational work to be found in many scientific codes. For some codes, almost all of the computational intensity can be found in the use of a library for, e.g., eigenvalue solution or FFTs. If such a code makes use of libraries for their solution, portability is very often assured. Indeed, even if a particular library has not been ported to a new architecture at a given time, the library source code is often available and can be compiled by the user on the new platform. However, the best performance is realized when either the library maintainers or the machine vendor (or both) undertake development to optimize a given library on a particular platform. This obvious advantage has been realized by vendors, and for many of the libraries referred to earlier, this optimization is done as a matter of course.

Conversely, performance cannot be guaranteed with the same degree of certainty. First, although libraries often do encapsulate a good measure of the required work, in most cases this is not all of the work, including what is often strictly serial work. This fundamental constraint is sometimes exacerbated by the fact that architecture-specific implementations are evolving, despite the best efforts of both vendors and library maintainers.

Codes with obvious "hot spots" can often make immediate use of libraries to achieve performance portability. This is often easiest for codes written in Fortran and C, whereas bindings to many libraries in C++ can be lacking or somewhat arcane to use. One of the biggest concerns in using libraries for extant codes is the frequent requirement to recast data structures used in the code to the format used by the library. The best approach to ameliorate this problem is often to simply use a memory copy: The relative cost of the copy compared to the work done in the library is often small, and the use of a localized copy obviates the need to change data structures pervasively throughout the code.

  • BLAS/LAPACK - dense linear algebra

    • BLAS and LAPACK are often contained in vendor-supplied library collections, like:

    • In addition, other platform-specific implementations are available, like:

  • FFTW - Fast Fourier Transform

    • Like LAPACK/BLAS, FFTW-like APIs can be found in MKL and ACML
  • PETSc - PDE solvers

    • PETSc is much more like a framework, and often requires more extensive code changes to use efficiently

Benefits and Challenges

Benefits

  • Often encapsulate much of the computational intensity found in scientific codes
  • Can allow immediate portability under some circumstances
  • Performance becomes a task for library authors/maintainers

Challenges

  • Limited set of portable libraries at present
  • May not capture all the important/expensive tasks in a given code
  • Often require recasting data structures to match library requirements
  • Opaque interior threading models