Academic Journal
Προβολή σε EDS

NeutronJittor: Efficient CPU-GPU memory management for GNN training with Jittor.

Λεπτομέρειες βιβλιογραφικής εγγραφής
Τίτλος: NeutronJittor: Efficient CPU-GPU memory management for GNN training with Jittor.
Συγγραφείς: Lu, Shizhan1, Fu, Zhenbo1, Wang, Qiange1, Zhang, Yanfeng1 zhangyf@mail.neu.edu.cn, Xu, Yuehao1, Ai, Xin1, Wei, Zhewei2, Yu, Ge1
Πηγή: Fundamental Research. Jul2026, Vol. 6 Issue 4, p2602-2611. 10p.
Θεματικοί όροι: *Graph neural networks, *Computer memory management, *Heterogeneous computing, *Machine learning, *Process optimization, *Graphics processing units
Περίληψη: Graph Neural Networks (GNNs) have been widely adopted across domains, and full-graph training is often preferred for its accuracy and stability. However, this approach demands large memory footprints, requiring dynamic memory management during GPU execution. The highly irregular data access of graphs makes manual management impractical, motivating the need for user-transparent memory management. We observe that existing user-transparent memory management frameworks are inefficient when applied to GNN workloads. Their inefficiency stems from two fundamental sources: redundant computation caused by repeatedly processing shared vertices across subgraphs, and frequent CPU–GPU communication caused by suboptimal data movement strategies. To overcome these limitations, we introduce NeutronJittor, an efficient GNN training system with user-transparent memory management. Firstly, we design an embedding-reuse execution strategy that reuses the intermediate embeddings of shared vertices within an epoch, allowing later subgraphs to skip redundant computations and release temporary data sooner, reducing redundant computation and lowering GPU memory pressure. Secondly, we develop a spatial–temporal-aware data scheduler (ST-Aware scheduler) that ranks vertices by structural and temporal priority, enabling the system to decide which data should stay on GPU and reducing unnecessary CPU–GPU data movement. Extensive experiments on real-world datasets with GCN and GAT models show that NeutronJittorreduces communication volume by 65.7%–94.3% and achieves 3.71×–11.61× speedup compared to state-of-the-art systems (DGL and NeutronStar), while maintaining comparable model accuracy. [ABSTRACT FROM AUTHOR]
Βάση Δεδομένων: Academic Search Index
FullText Text:
  Availability: 0
CustomLinks:
  – Url: https://www.doi.org/10.1016/j.fmre.2026.04.035?
    Name: ScienceDirect (all content) (s7799221)
    Category: fullText
    Text: View record from ScienceDirect
    MouseOverText: View record from ScienceDirect
  – Url: https://resolver.ebsco.com/c/fiv2js/result?sid=EBSCO:asx&genre=article&issn=20969457&ISBN=&volume=6&issue=4&date=20260701&spage=2602&pages=2602-2611&title=Fundamental Research&atitle=NeutronJittor%3A%20Efficient%20CPU-GPU%20memory%20management%20for%20GNN%20training%20with%20Jittor.&aulast=Lu%2C%20Shizhan&id=DOI:10.1016/j.fmre.2026.04.035
    Name: Full Text Finder (for New FTF UI) (ns324271)
    Category: fullText
    Text: Full Text Finder
    MouseOverText: Full Text Finder
Header DbId: asx
DbLabel: Academic Search Index
An: 196279227
RelevancyScore: 1431
AccessLevel: 6
PubType: Academic Journal
PubTypeId: academicJournal
PreciseRelevancyScore: 1430.76818847656
IllustrationInfo
Items – Name: Title
  Label: Title
  Group: Ti
  Data: NeutronJittor: Efficient CPU-GPU memory management for GNN training with Jittor.
– Name: Author
  Label: Authors
  Group: Au
  Data: <searchLink fieldCode="AR" term="%22Lu%2C+Shizhan%22">Lu, Shizhan</searchLink><relatesTo>1</relatesTo><br /><searchLink fieldCode="AR" term="%22Fu%2C+Zhenbo%22">Fu, Zhenbo</searchLink><relatesTo>1</relatesTo><br /><searchLink fieldCode="AR" term="%22Wang%2C+Qiange%22">Wang, Qiange</searchLink><relatesTo>1</relatesTo><br /><searchLink fieldCode="AR" term="%22Zhang%2C+Yanfeng%22">Zhang, Yanfeng</searchLink><relatesTo>1</relatesTo><i> zhangyf@mail.neu.edu.cn</i><br /><searchLink fieldCode="AR" term="%22Xu%2C+Yuehao%22">Xu, Yuehao</searchLink><relatesTo>1</relatesTo><br /><searchLink fieldCode="AR" term="%22Ai%2C+Xin%22">Ai, Xin</searchLink><relatesTo>1</relatesTo><br /><searchLink fieldCode="AR" term="%22Wei%2C+Zhewei%22">Wei, Zhewei</searchLink><relatesTo>2</relatesTo><br /><searchLink fieldCode="AR" term="%22Yu%2C+Ge%22">Yu, Ge</searchLink><relatesTo>1</relatesTo>
– Name: TitleSource
  Label: Source
  Group: Src
  Data: <searchLink fieldCode="JN" term="%22Fundamental+Research%22">Fundamental Research</searchLink>. Jul2026, Vol. 6 Issue 4, p2602-2611. 10p.
– Name: Subject
  Label: Subject Terms
  Group: Su
  Data: *<searchLink fieldCode="DE" term="%22Graph+neural+networks%22">Graph neural networks</searchLink><br />*<searchLink fieldCode="DE" term="%22Computer+memory+management%22">Computer memory management</searchLink><br />*<searchLink fieldCode="DE" term="%22Heterogeneous+computing%22">Heterogeneous computing</searchLink><br />*<searchLink fieldCode="DE" term="%22Machine+learning%22">Machine learning</searchLink><br />*<searchLink fieldCode="DE" term="%22Process+optimization%22">Process optimization</searchLink><br />*<searchLink fieldCode="DE" term="%22Graphics+processing+units%22">Graphics processing units</searchLink>
– Name: Abstract
  Label: Abstract
  Group: Ab
  Data: Graph Neural Networks (GNNs) have been widely adopted across domains, and full-graph training is often preferred for its accuracy and stability. However, this approach demands large memory footprints, requiring dynamic memory management during GPU execution. The highly irregular data access of graphs makes manual management impractical, motivating the need for user-transparent memory management. We observe that existing user-transparent memory management frameworks are inefficient when applied to GNN workloads. Their inefficiency stems from two fundamental sources: redundant computation caused by repeatedly processing shared vertices across subgraphs, and frequent CPU–GPU communication caused by suboptimal data movement strategies. To overcome these limitations, we introduce NeutronJittor, an efficient GNN training system with user-transparent memory management. Firstly, we design an embedding-reuse execution strategy that reuses the intermediate embeddings of shared vertices within an epoch, allowing later subgraphs to skip redundant computations and release temporary data sooner, reducing redundant computation and lowering GPU memory pressure. Secondly, we develop a spatial–temporal-aware data scheduler (ST-Aware scheduler) that ranks vertices by structural and temporal priority, enabling the system to decide which data should stay on GPU and reducing unnecessary CPU–GPU data movement. Extensive experiments on real-world datasets with GCN and GAT models show that NeutronJittorreduces communication volume by 65.7%–94.3% and achieves 3.71×–11.61× speedup compared to state-of-the-art systems (DGL and NeutronStar), while maintaining comparable model accuracy. [ABSTRACT FROM AUTHOR]
PLink https://search.ebscohost.com/login.aspx?direct=true&site=eds-live&db=asx&AN=196279227
RecordInfo BibRecord:
  BibEntity:
    Identifiers:
      – Type: doi
        Value: 10.1016/j.fmre.2026.04.035
    Languages:
      – Code: eng
        Text: English
    PhysicalDescription:
      Pagination:
        PageCount: 10
        StartPage: 2602
    Subjects:
      – SubjectFull: Graph neural networks
        Type: general
      – SubjectFull: Computer memory management
        Type: general
      – SubjectFull: Heterogeneous computing
        Type: general
      – SubjectFull: Machine learning
        Type: general
      – SubjectFull: Process optimization
        Type: general
      – SubjectFull: Graphics processing units
        Type: general
    Titles:
      – TitleFull: NeutronJittor: Efficient CPU-GPU memory management for GNN training with Jittor.
        Type: main
  BibRelationships:
    HasContributorRelationships:
      – PersonEntity:
          Name:
            NameFull: Lu, Shizhan
      – PersonEntity:
          Name:
            NameFull: Fu, Zhenbo
      – PersonEntity:
          Name:
            NameFull: Wang, Qiange
      – PersonEntity:
          Name:
            NameFull: Zhang, Yanfeng
      – PersonEntity:
          Name:
            NameFull: Xu, Yuehao
      – PersonEntity:
          Name:
            NameFull: Ai, Xin
      – PersonEntity:
          Name:
            NameFull: Wei, Zhewei
      – PersonEntity:
          Name:
            NameFull: Yu, Ge
    IsPartOfRelationships:
      – BibEntity:
          Dates:
            – D: 01
              M: 07
              Text: Jul2026
              Type: published
              Y: 2026
          Identifiers:
            – Type: issn-print
              Value: 20969457
          Numbering:
            – Type: volume
              Value: 6
            – Type: issue
              Value: 4
          Titles:
            – TitleFull: Fundamental Research
              Type: main
ResultId 1