The era of big data creates opportunities for carrying out scientific simulations at exascale.With increasing data size, the complexity of the design and execution of scientific applications demand the use of high-level tools, namely workflow systems on supercomputers. The performance of workflow systems has paramount importance since the goal of running scientific applications on a supercomputing system is to improve the end-to-end performance. The first part of this dissertation focuses on developing algorithms for intergroup All-to-All broadcast communication, which is an important data exchange pattern between workflow components. The second part of this thesis focuses on improving the two-phase I/O algorithm, which allows faster parallel file I/O operations by workflow components. The rest of this thesis contains miscellaneous works that apply machine learning approaches for improving applications’ prediction accuracy and execution performance.

Creator

• Kang, Qiao

DOI

• https://doi.org/10.21985/n2-qx6a-9z80

Subject

• Computer engineering

Language

• en

Alternate Identifier

• etdadmin_upload_780039
• http://dissertations.umi.com/northwestern:15382

Keyword

• High performance computing
• Parallel Computing
• Parallel I/O

Date created

• 2020-01-01

Resource type

• Dissertation

Rights statement

• In Copyright