NESL: A Nested Data-Parallel Language

NESL is a parallel language designed to express parallel algorithms
and applications in a concise and clear manner, but yet achieve
reasonable performance on a variety of parallel machines.  The main
two ideas behind the language are the support for nested data-parallel
constructs, and a performance model based on keeping track of the work
and depth of a computation.  The nested data-parallel constructs allow
one to easily express algorithms that use either parallel
divide-and-conquer or nested parallel loops (these are hard to express
in most existing data-parallel languages).  The performance model
makes it possible to guarantee asymptotic bounds on running time on
various machine models.

In this talk I will give an overview of NESL and go through some
examples of algorithms, showing how they use nested parallelism and
how performance is analyzed.  I will also discuss the current
implementation, which runs on the Cray C90, Connection Machine CM-5,
IBM SP-2, and Intel Paragon.  NESL is a call-by-value functional
language and supports many of the features of standard functional
languages, such as polymorphic types, higher order functions, and a
limited form of Haskell style overloading.  The compiler uses
aggressive type specialization to achieve good performance on
polymorphic code.