Author: Benedikt Bleimhofer
Title: Hierarchical Arrays for Efficient and Productive Data-Intensive Parallel Computing
Abstract:
Today's high performance computing (HPC) systems are characterized by
massive parallelism and multi-level memory hierarchies. Currently most
parallel programs are written using either multithreading (e.g.,
OpenMP) or message passing (e.g., MPI). While threads are easier to
program than message passing and provide implicit communication via
shared memory, their effectiveness is limited on large systems.
Message passing on the other hand scales well on current clusters and
provides explicit locality control. However, the programming is very
complex and error-prone due to the explicit communication. Partitioned
Global Address Space (PGAS) approaches are a relatively new way to
combine the advantages of the two aforementioned models. Even though
some of the most popular PGAS languages have already been introduced
in 2004, the knowledge about these new programming approaches and
their acceptance is generally very low. We believe that this is mainly
due to the fact that most current PGAS approaches are implemented as
new programming languages and thus require a complete rewriting of
existing application code. Furthermore, most current PGAS languages do
not support multilevel memory hierarchies and do thus not match
current hardware. In this thesis Hierarchical Arrays (HA), a
data-structure oriented C++ template library, providing hierarchical
PGAS-like abstractions for essential data types like multidimensional
arrays and distributed lists is designed and prototypically
implemented. The general principles for the design are the explicit
control of the hierarchical data layout by the programmer, the use of
one-sided message primitives, and the ability of the library to
co-exist with current parallel programming models such as MPI, in
order to support incremental adoption.