Mercator: A GPGPU framework for irregular streaming applications

Stephen V. Cole, Jeremy Buhler

Research output: Chapter in Book/Report/Conference proceedingConference contributionpeer-review

9 Scopus citations

Abstract

GPUs have a natural affinity for streaming applications exhibiting consistent, predictable dataflow. However, many high-impact irregular streaming applications, including sequence pattern matching, decision-tree and decision-cascade evaluation, and large-scale graph processing, exhibit unpredictable dataflow due to data-dependent filtering or expansion of the data stream. Existing GPU frameworks do not support arbitrary irregular streaming dataflow tasks, and developing application-specific optimized implementations for such tasks requires expert GPU knowledge. We introduce MERCATOR, a lightweight framework supporting modular CUDA streaming application development for irregular applications. A developer can use MERCATOR to decompose an irregular application for the GPU without explicitly remapping work to threads at runtime. MERCATOR applications are efficiently parallelized on the GPU through a combination of replication across blocks and queueing between nodes to accommodate irregularity. We quantify the performance impact of MERCATOR's support for irregularity and illustrate its utility by implementing a biological sequence comparison pipeline similar to the well-known NCBI BLASTN algorithm. MERCATOR code is available by request to the first author.

Original languageEnglish
Title of host publicationProceedings - 2017 International Conference on High Performance Computing and Simulation, HPCS 2017
EditorsWaleed W. Smari
PublisherInstitute of Electrical and Electronics Engineers Inc.
Pages727-736
Number of pages10
ISBN (Electronic)9781538632505
DOIs
StatePublished - Sep 12 2017
Event15th International Conference on High Performance Computing and Simulation, HPCS 2017 - Genoa, Italy
Duration: Jul 17 2017Jul 21 2017

Publication series

NameProceedings - 2017 International Conference on High Performance Computing and Simulation, HPCS 2017

Conference

Conference15th International Conference on High Performance Computing and Simulation, HPCS 2017
Country/TerritoryItaly
CityGenoa
Period07/17/1707/21/17

Keywords

  • GPU
  • Irregular computation
  • Parallel computing
  • SIMD
  • Streaming dataflow

Fingerprint

Dive into the research topics of 'Mercator: A GPGPU framework for irregular streaming applications'. Together they form a unique fingerprint.

Cite this