BifurKTM: Approximately consistent distributed transactional memory for GPUs

Samuel Irving, Lu Peng, Costas Busch, Jih Kwon Peir

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Abstract

We present BifurKTM, the first read-optimized Distributed Transactional Memory system for GPU clusters. The BifurKTM design includes: GPU KoSTM, a new software transactional memory conflict detection scheme that exploits relaxed consistency to increase throughput; and KoDTM, a Distributed Transactional Memory model that combines the Data- and Control- flow models to greatly reduce communication overheads. Despite the allure of huge speedups, GPUs are limited in use due to their programmability and extreme sensitivity to workload characteristics. These become daunting concerns when considering a distributed GPU cluster, wherein a programmer must design algorithms to hide communication latency by exploiting data regularity, high compute intensity, etc. The BifurKTM design allows GPU programmers to exploit a new workload characteristic: the percentage of the workload that is Read-Only (e.g. reads but does not modify shared memory), even when this percentage is not known in advance. Programmers designate transactions that are suitable for Approximate Consistency, in which transactions “appear” to execute at the most convenient time for preventing conflicts. By leveraging Approximate Consistency for Read-Only transactions, the BifurKTM runtime system offers improved performance, application flexibility, and programmability without introducing any errors into shared memory. Our experiments show that Approximate Consistency can improve BkTM performance by up to 34x in applications with moderate network communication utilization and a read-intensive workload. Using Approximate Consistency, BkTM can reduce GPU-to-GPU network communication by 99%, reduce the number of aborts by up to 100%, and achieve an average speedup of 18x over a similarly sized CPU cluster while requiring minimal effort from the programmer.

Original languageEnglish (US)
Title of host publication12th Workshop on Parallel Programming and Run-Time Management Techniques for Many-Core Architectures and 10th Workshop on Design Tools and Architectures for Multicore Embedded Computing Platforms, PARMA-DITAM 2021
EditorsJoao Bispo, Stefano Cherubin, Jose Flich
PublisherSchloss Dagstuhl- Leibniz-Zentrum fur Informatik GmbH, Dagstuhl Publishing
ISBN (Electronic)9783959771818
DOIs
StatePublished - Mar 1 2021
Event12th Workshop on Parallel Programming and Run-Time Management Techniques for Many-Core Architectures and 10th Workshop on Design Tools and Architectures for Multicore Embedded Computing Platforms, PARMA-DITAM 2021 - Budapest, Hungary
Duration: Jan 19 2021 → …

Publication series

NameOpenAccess Series in Informatics
Volume88
ISSN (Print)2190-6807

Conference

Conference12th Workshop on Parallel Programming and Run-Time Management Techniques for Many-Core Architectures and 10th Workshop on Design Tools and Architectures for Multicore Embedded Computing Platforms, PARMA-DITAM 2021
Country/TerritoryHungary
CityBudapest
Period1/19/21 → …

Keywords

  • Approximate Consistency
  • Distributed Transactional Memory
  • GPU

ASJC Scopus subject areas

  • Geography, Planning and Development
  • Modeling and Simulation

Fingerprint

Dive into the research topics of 'BifurKTM: Approximately consistent distributed transactional memory for GPUs'. Together they form a unique fingerprint.

Cite this