13270

Legion: Programming Distributed Heterogeneous Architectures with Logical Regions

Michael Edward Bauer
Stanford University
Stanford University, 2014

@phdthesis{bauer2014legion,

   title={LEGION: PROGRAMMING DISTRIBUTED HETEROGENEOUS ARCHITECTURES WITH LOGICAL REGIONS},

   author={Bauer, Michael Edward},

   year={2014},

   school={STANFORD UNIVERSITY}

}

Download Download (PDF)   View View   Source Source   Source codes Source codes

Package:

850

views

This thesis covers the design and implementation of Legion, a new programming model and runtime system for targeting distributed heterogeneous machine architectures. Legion introduces logical regions as a new abstraction for describing the structure and usage of program data. We describe how logical regions provide a mechanism for applications to express important properties of program data, such as locality and independence, that are often ignored by current programming systems. We also show how logical regions allow programmers to scope the usage of program data by different computations. The explicit nature of logical regions makes these properties of programs manifest, allowing many of the challenging burdens of parallel programming, including dependence analysis and data movement, to be off-loaded from the programmer to the programming system. Logical regions also improve the programmability and portability of applications by decoupling the specification of a program from how it is mapped onto a target architecture. Logical regions abstractly describe sets of program data without requiring any specification regarding the placement or layout of data. To control decisions about the placement of computations and data, we introduce a novel mapping interface that gives an application programmatic control over mapping decisions at runtime. Different implementations of the mapper interface can be used to port applications to new architectures and to explore alternative mapping choices. Legion guarantees that the decisions made through the mapping interface are independent of the correctness of the program, thus facilitating easy porting and tuning of applications to new architectures with different performance characteristics. Using the information provided by logical regions, an implementation of Legion can automatically extract parallelism, manage data movement, and infer synchronization. We describe the algorithms and data structures necessary for efficiently performing these operations. We further show how the Legion runtime can be generalized to operate as a distributed system, making it possible for Legion applications to scale well. As both applications and machines continue to become more complex, the ability of Legion to relieve application developers of many of the tedious responsibilities they currently face will become increasingly important. To demonstrate the performance of Legion, we port a production combustion simulation, called S3D, to Legion. We describe how S3D is implemented within the Legion programming model as well as the different mapping strategies that are employed to tune S3D for runs on different architectures. Our performance results show that a version of S3D running on Legion is nearly three times as fast as comparable state-of-the-art versions of S3D when run at 8192 nodes on the number two supercomputer in the world.
No votes yet.
Please wait...

* * *

* * *

HGPU group © 2010-2017 hgpu.org

All rights belong to the respective authors

Contact us: