PERFORMANCE

========================
Chapel Performance Notes
========================

Though Chapel has been designed to ultimately yield high performance,
our focus to date has predominantly been on implementing its features
correctly and providing user-supported control of features like array
implementations, loop schedules, and architectural descriptions.  To
that end, the current compiler is lacking several key optimizations
and therefore is often not competitive with hand-coded C, Fortran,
MPI, and the like.  We are currently working on improving performance
to close the gap between Chapel and conventional compiled languages
and parallel programming models.

In measuring Chapel performance, users should be sure to compile with
the --fast flag once their program is correct.  This is a meta-flag
that turns on C-level optimizations and turns off several
execution-time correctness checks (bounds checks, NULL pointer checks,
etc.).  See the 'chpl' man page for more information.

To characterize Chapel performance, generally speaking...

* single-locale (CHPL_COMM=none | --local) compilations perform better
  than multi-locale (CHPL_COMM!=none | --no-local) compilations;

* 1D loops/arrays perform better than multidimensional cases;

* codes with structured communication (e.g., stencils) tend not to
  perform competitively with hand-coded computations, whereas
  embarrassingly parallel and unstructured communications tend to be
  more competitive.

We are currently working to improve Chapel performance with each
release and are making significant strides.  To track our progress
over time, refer to:

        http://chapel.sourceforge.net/perf/

(particularly the release-over-release performance graphs, which
provide a way to track performance trends on a release-by-release
granularity).