Skip to content

Partitioned Global Address Space

A programming model makes eligible data referable across locality domains while exposing each location's partition or affinity and defining local and remote access.

Version
v1 · 2026-10-17 · History
Domain-specific #
14244
Domain group
Applied Sciences & Engineering
Origin domain
Computer Science & Software Engineering
Subdomain
Parallel Programming → Computer Science & Software Engineering
Aliases
PGAS

Core Idea

A partitioned global address space (PGAS) is a parallel-programming model making eligible data referable across localities while exposing its partition or affinity. A language or library defines access semantics. “Global” does not expose every private datum, and “partitioned” does not equalize remote and local costs.[ref-46b401224f1e][ref-9ed7f8a06970]

Scope of Application

The two documented habitats use different interfaces with eligible cross-locality data and visible placement:

  • Chapel language. A worked nested on example updates x from Locale 2 while x remains on Locale 1; on alone creates no parallel task.[^ref-46b401224f1e]
  • OpenSHMEM library. Listing 6 puts five selected values from PE 0's private source into PE 1's symmetric target, with barriers around observation.[^ref-9ed7f8a06970]

Both are worked examples, not observed deployments or performance measurements. A locale is one unit, not the whole model.[ref-46b401224f1e][ref-9ed7f8a06970]

Clarity

Ask which data are remotely eligible, where they reside, and which operation makes access valid. OpenSHMEM's private source and symmetric target differ; Chapel's remote update leaves x at home. A global reference surface alone promises neither uniform cost, automatic concurrency, nor visibility without the contract and synchronization.[ref-46b401224f1e][ref-9ed7f8a06970]

Manages Complexity

PGAS lets programmers refer to eligible distributed data without expressing every access as a private message, while exposing locality for placement. Its five roles are locality domains, eligible data, visible affinity, cross-domain access, and operation semantics. Compact syntax may still entail communication; the cited systems establish no universal speedup or consistency rule.[ref-46b401224f1e][ref-9ed7f8a06970]

Abstract Reasoning

Identify the eligible object, target affinity, access operation, and synchronization before judging legality or cost. Chapel code on Locale 2 reaches x on Locale 1; OpenSHMEM names the destination PE and symmetric target. Listing 6's count argument selects five elements despite a conflicting “ten words” comment. Neither example makes every datum remotely accessible.[ref-46b401224f1e][ref-9ed7f8a06970]

Knowledge Transfer

The roles transfer between Chapel's language and OpenSHMEM's library, but syntax and memory-order rules differ. Interface is an internal constituent: each PGAS model declares eligible objects, visible locality, operations, and effects through a bounded contract. A hardware locality or ordinary cross-boundary message alone lacks the full model.[ref-46b401224f1e][ref-9ed7f8a06970]

Relationships to Other Abstractions

Local relationship map for Partitioned Global Address SpaceParents appear above the current abstraction, mutual partners to the right, and children below. Node labels state whether each abstraction is prime or domain-specific; colors identify relation types.Partitioned GlobalAddress SpaceDOMAINPrime abstraction: Interface — is part ofInterfacePRIME

Current abstraction Partitioned Global Address Space Domain-specific

Parents (1) — more general patterns this builds on

  • Partitioned Global Address Space is part of Interface Prime

    Every PGAS programming model contains a bounded access interface that declares eligible remote data, visible locality, operations, and their effects.

Hierarchy path (1) — routes to 1 parentless root

Neighborhood in Abstraction Space

Partitioned Global Address Space sits in a moderately populated region (57th percentile for distinctiveness): it has near-neighbors but no dense thicket of look-alikes.

Family — Network Security Vulnerabilities & Trust (30 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-17

References

[^ref-46b401224f1e]: Steven J. Deitz, Bradford L. Chamberlain, Samuel Figueroa, and David Iten, “HPCC STREAM and RA in Chapel: Performance and Potential,” CUG 2009 Proceedings, HPCC STREAM and RA in Chapel: Performance and Potential. Consulted §2.1.2 locale query; §2.2 nested on x, PDF pp. 2–3; §2.3 distinction from task parallelism, PDF p. 3; §§3–4 locality/distribution implementation, PDF pp. 3–7. This is original author design and worked code, not reviewer-executed application evidence. [^ref-9ed7f8a06970]: OpenSHMEM (project authors and collaborators), OpenSHMEM Application Programming Interface Version 1.0 FINAL, 2012. Consulted §§2.1–2.2 PGAS and SHMEM, printed pp. 6–7; §§5–6 memory/execution; §8.8 strided puts; Appendix E.1 Listing 6, printed p. 108 / zero-based PDF p. 107. The listing is adapted from SGI man pages and was not reviewer-executed.