Visualising DMA operations for improved parallel performance

Research output: Contribution to conferencePresentation

Abstract

A single-source heterogeneous parallel programming model may be attractive for a number of reasons, including: the provision of an intuitive development interface; ease of debugging; and the facilitation of a concise codebase. Common examples are CUDA; OpenACC; and C++ AMP. At Codeplay Software we also provide a range of single-source C++ solutions within our Offload technology.

Such programming models ensure that large, existing codebases are rapidly parallelised for heterogeneous architectures. A second edge to the sword exists though, in the consequent implicit nature of data transfers occurring between distinct physical memories and corresponding address spaces.

Our Offload technology represents a novel, patented, set of heterogeneous parallel programming language technologies, featuring the integration of enhanced pointer types which are implicitly, statically, and immutably associated with a single address space; and typically correspond to a unit of physical memory.

In this talk I will describe our ongoing research work within the EU Framework 7 project LPGPU to visualise and react to expensive direct memory access operations, by applying new tools which augment the Offload C++ compiler for PlayStation 3.
Original languageEnglish
Publication statusPublished - 8 May 2013
Externally publishedYes
Event2013 MMnet Workshop
: Language and Runtime Support for Concurrent Systems
- Heriot-Watt University, Edinburgh, United Kingdom
Duration: 8 May 2013 → …
http://www.macs.hw.ac.uk/~hwloidl/Events/MMNet/mmnet13.html

Workshop

Workshop2013 MMnet Workshop
Abbreviated titleMMnet'13
Country/TerritoryUnited Kingdom
CityEdinburgh
Period8/05/13 → …
Internet address

Keywords

  • Memory
  • C++
  • Parallelism
  • Visualisation

Fingerprint

Dive into the research topics of 'Visualising DMA operations for improved parallel performance'. Together they form a unique fingerprint.

Cite this