Adaptive runtime techniques for node-aware resource optimization

Chandrasekar, Kavitha

Permalink

https://hdl.handle.net/2142/129286 Copy

Description

Title

Adaptive runtime techniques for node-aware resource optimization

Author(s)

Chandrasekar, Kavitha

Issue Date

2025-04-30

Director of Research (if dissertation) or Advisor (if thesis)

Kale, Laxmikant

Doctoral Committee Chair(s)

Kale, Laxmikant

Committee Member(s)

• Torrellas, Josep
• Amato, Nancy
• Rountree, Barry

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• High Performance Computing
• Node-aware resource tuning
• Message aggregation
• Load balancing

Abstract

Adaptive runtime systems have to deal with an increasing number of cores per node. Adaptive algorithms that worked well when the number of cores per node was smaller now run into many challenges. This dissertation is aimed at addressing these challenges along multiple frontiers. The first challenge is simply about how many cores on each node to utilize for a given application. We analyzed this issue from the point of view of performance as well as energy saving, and demonstrate adaptive runtime techniques for tuning the number of cores used during execution. Secondly, many of the issues discussed here are meaningful mainly when one is using a shared-memory process spanning multiple cores of a node, rather than a "Charm++ everywhere" (analogous to MPI everywhere) mode, where each process runs on a core of its own. The latter is fundamentally sub-optimal, because it precludes resource and data-structure sharing. However, the former suffers from performance issues, especially in communication, because it is harder to expose and exploit parallelism in the network interface usage. We address this issue with comprehensive analysis and communication layer design driven by such analysis to demonstrate on-par communication performance and/or identify additional issues that need to be dealt with. Thirdly, adaptive runtime systems often have to deal with fine-grained communication. This necessitates adaptive aggregation of many short messages. However, large multicore nodes complicate this aggregation functionality, along with the diverse needs of applications. We develop metrics for characterizing the use-cases, and develop adaptive algorithms that work well on multicore nodes and for varied use cases. Fourthly, dynamic load balancing is another challenge: migrating work units to cores to effectively balance load in dynamic applications needs to take into account the multi-level nature of communication, while also addressing scalability issues created by the huge number of cores on modern machines. We demonstrate node aware, and communication aware strategies, including refinement strategies that aim at minimal migration and low load balancing overhead.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129286

Copyright and License Information

Copyright 2025 Kavitha Chandrasekar

Owning Collections