The Fortress Language Specification - CiteSeerX

Chapter 32
Parallelism and Locality
Fortress is designed to make parallel programming as simple and as painless as possible. This chapter describes the
internals of Fortress parallelism designed for use by authors of library code (such as distributions, generators, and
arrays). We adopt a multi-tiered approach to parallelism:
• At the highest level, we provide libraries which allocate locality-aware distributed arrays (Section 32.2) and
implicitly parallel constructs such as tuples and loops. Synchronization is accomplished through the use of
atomic sections (Section 13.23). More complex synchronization makes use of abortable atomicity, described in
Section 32.3.
• There is an extensive library of distributions, which permits the programmer to specify locality and data distribution
explicitly (Section 32.5).
• Immediately below that, the at expression requests that a computation take place in a particular region of the
machine (Section 32.7). We also provide a mechanism to terminate a spawned thread early (Section 32.6).
• Finally, there are mechanisms for constructing new generators via recursive subdivision into tree structures
with individual elements at the leaves. Section 32.8 explains how iterative constructs such as for loops and
comprehensions are desugared into calls to methods of trait Generator, and how new instances of this trait may
be defined.
We begin by describing the abstraction of regions, which Fortress uses to describe the machine on which a program is
run.
32.1 Regions
Every thread and every object in Fortress, and every element of a Fortress array, has an associated region. The region
in which an object o resides can be obtained by calling o.region . Regions abstractly describe the structure of the
machine on which a Fortress program is running. They are organized hierarchically to form a tree, the region hierarchy,
reflecting in an abstract way the degree of locality which those regions share. The different levels of this tree reflect
underlying machine structure, such as execution engines within a CPU, memory shared by a group of processors, or
resources distributed across the entire machine. Objects which reside in regions near the leaves of the tree are local
entities; those which reside at higher levels of the region tree are logically spread out. The method call r.isLocalTo(s)
returns true if r is contained within the region tree rooted at s .
It is important to understand that regions and the structures (such as distributions, Section 32.5) built on top of them
exist purely for performance purposes. The placement of a thread or an object does not have any semantic effect on
215