Skip to main content
Awards and Recognition | Mathematics and Computer Science

Argonne researchers to present latest work at high-performance computing conference

Researchers in Argonne’s Mathematics and Computer Science (MCS) division will present five papers at HiPC 2023, the IEEE International Conference on High Performance Computing, Data, and Analytics.

This year’s meeting is divided into two focus areas: high-performance computing and scalable data science. The data reduction and management team in Argonne’s MCS division had five papers accepted in the first focus area.

HiPC is a major international conference, and having five papers accepted out of a total of 42 is impressive,” said Franck Cappello, a senior scientist in Argonne’s MCS division.

Three of the papers accepted for the conferences deal with compression issues (boldface means an MCS division staff member).

  • Arham Khan, Sheng Di, Kyle Chard, Ian Foster, Franck Cappello,” SECRE: Surrogate-based Error-controlled Lossy Compression Ratio Estimation Framework.”  This paper proposes a compression ratio estimation method for four state-of-the-art error-controlled lossy compressors by devising a lightweight compression surrogate for each. Experiments with four real-world simulation datasets show that SECRE can obtain highly accurate compression ratio estimates with low execution overhead.
  • Pu Jiao, Sheng Di, Jinyang Liu, Xin Liang, Franck Cappello, Characterization and Detection of Artifacts for Error-controlled Lossy Compressors.”  Error-bounded lossy compressors may have serious artifact issues in situations with relatively large error bound or high compression ratios. In this paper the authors characterize the artifacts for multiple state-of-the-art error-bounded lossy compressors, provide an in-depth analysis of the root cause of these artifacts, and develop an efficient artifact detection algorithm for each type of artifact.
  • Avinash Maurya, Bogdan Nicolae, M. Mustafa Rafique, Franck Cappello, Towards Efficient I/O Pipelines using Accumulated Compression.” The authors present strategies to optimize the trade-off between compressing checkpoints instantly and enqueuing transfers immediately vs accumulating snapshots and delaying compression to achieve faster compression throughput.

Two other papers focus on deep learning:

  • Kevin Assogba, Mustafa Rafique, Bogdan Nicolae, Optimizing the Training of Co-Located Deep Learning Models Using Cache-Aware Staggering.”  How can one efficiently train co-located deep learning models that share the same input data stored initially on a remote repository such as a parallel file system? The authors propose a performance model to find the optimal staggering that produces the minimum makespan, which also reduces resource utilization.
  • Robert Underwood, Meghana Madhastha, Randal Burns, Bogdan Nicolae, Understanding Patterns of Deep Learning Model Evolution in Network Architecture Search.” Network architecture search (NAS) is a foundational method for identifying viable deep learning model architectures. NAS, however, is computationally and resource intensive. In this paper the researchers show how the evolution of the model structure is influenced by the regularized evolution algorithm. The work is a step toward improving the scalability and performance of network architecture search using regularized evolution and other genetic algorithms.

This year’s HiPC conference will take place in Goa, India, December 18-21, 2023. For more information, see the conference website at https://​hipc​.org/.