Regent : a high-productivity programming language for implicit parallelism with logical regions
Abstract/Contents
- Abstract
- Modern supercomputers are dominated by distributed-memory machines. State of the art high-performance scientific applications targeting these machines are typically written in low-level, explicitly parallel programming models that enable maximal performance but expose the user to programming hazards such as data races and deadlocks. Conversely, implicitly parallel models isolate the user from these hazards by providing easy-to-use sequential semantics and place responsibility for parallelism and data movement on the system. However, traditional implementations of implicit parallelism suffer from substantial limitations: static, compiler-based implementations restrict the programming model to exclude dynamic features needed for unstructured applications, while dynamic, runtime-based approaches suffer from a sequential bottleneck that limits the scalability of the system. We present Regent, a programming language designed to enable a hybrid static and dynamic analysis of implicit parallelism. Regent programs are composed of tasks (functions with annotated data usage). Program data is stored in regions (hierarchical collections); regions are dynamic, first-class values, but are named statically in the type system to ensure correct usage and analyzability of programs. Tasks may execute in parallel when they are mutually independent as determined by the annotated usage (read, write, etc.) of regions passed as task arguments. A Regent implementation is responsible for automatically discovering parallelism in a Regent program by analyzing the executed tasks in program order. A naive implementation of Regent would suffer from a sequential bottleneck as tasks must be analyzed sequentially at runtime to discover parallelism, limiting scalability. We present an optimizing compiler for Regent which transforms implicitly parallel programs into efficient explicitly parallel code. By analyzing the region arguments to tasks, the compiler is able to determine the data movement implied by the sequence of task calls, even in the presence of unstructured and data-dependent application behavior. The compiler can then replace the implied data movement with explicit communication and synchronization for efficient execution on distributed-memory machines. We measure the performance and scalability of several Regent programs on large supercomputers and demonstrate that optimized Regent programs perform comparably to manually optimized explicitly parallel programs.
Description
| Type of resource | text |
|---|---|
| Form | electronic; electronic resource; remote |
| Extent | 1 online resource. |
| Publication date | 2017 |
| Issuance | monographic |
| Language | English |
Creators/Contributors
| Associated with | Slaughter, Elliott |
|---|---|
| Associated with | Stanford University, Computer Science Department. |
| Primary advisor | Aiken, Alexander |
| Thesis advisor | Aiken, Alexander |
| Thesis advisor | Levis, Philip |
| Thesis advisor | Olukotun, Oyekunle Ayinde |
| Advisor | Levis, Philip |
| Advisor | Olukotun, Oyekunle Ayinde |
Subjects
| Genre | Theses |
|---|
Bibliographic information
| Statement of responsibility | Elliott Slaughter. |
|---|---|
| Note | Submitted to the Department of Computer Science. |
| Thesis | Thesis (Ph.D.)--Stanford University, 2017. |
| Location | electronic resource |
Access conditions
- Copyright
- © 2017 by Elliott David Slaughter
- License
- This work is licensed under a Creative Commons Attribution Non Commercial 3.0 Unported license (CC BY-NC).
Also listed in
Loading usage metrics...