Design for Heterogeneity

Overview

Questions

• How should one design software for heterogeneous platforms

Objectives

• To understand how research software design needs to adapt for heterogeneous platforms

Mechanisms Needed by the Code

Underlying Ideas: Unification of Computational Expressions

Make the same code work on different devices

• Same algorithm different data layouts or operation sequence:
  • A way to let compiler know that ”this” expression can be specialized in many ways
  • Definition of specializations
  • Often done with template meta-programming
• More challenging if algorithms need to be fundamentally different
  • Support for alternatives

Underlying Ideas: Moving Work and Data to the Target

Parallelization Models

• Hierarchy in domain decomposition
  • Distributed memory model at node level – still very prevalent, likely to remain so for a while
  • Also done with PGAS models – shared with locality being important

Assigning work within the node

“Parallel For” or directives with unified memory

Directives or specific programming model for explicit data movement

More complex data orchestration system for asynchronous computation

Task based work distribution

Underlying Ideas: Mapping Work to Targets

This is how many abstraction layers work

• Infer the structure of the code

• Infer the map between algorithms and devices

• Infer the data movements

• Map computations to devices

These are specified either through constructs or pragmas

It can also be the end user who figures out the mapping

In either case performance depends upon how well the mapping is done

Features and Abstractions to be Added to Earlier Example

Mechanisms Needed by the Code : Example Flash-X

Construction of Application with Components and Tools

Key Points

• More separation of concerns and abstractions are needed