Rethinking communication libraries for asynchronous multithreaded runtimes

Yan, Jiakun

Permalink

https://hdl.handle.net/2142/132527 Copy

Description

Title

Rethinking communication libraries for asynchronous multithreaded runtimes

Author(s)

Yan, Jiakun

Issue Date

2025-11-25

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

Snir, Marc

Doctoral Committee Chair(s)

Snir, Marc

Committee Member(s)

• Gropp, William
• Kale, Laxmikant
• Yelick, Katherine
• Guo, Yanfei
• Kaiser, Hartmut

Department of Study

Siebel School Comp & Data Sci

Discipline

Computer Science

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• Communication library
• MPI
• Multithreaded message passing
• Task parallel
• Asynchronous many-task systems

Abstract

The evolution of architectures, algorithms, and programming models is driving communication toward greater asynchrony and concurrency, often in multithreaded environments. A representative example is the Asynchronous Many-Task (AMT) model, which emphasizes fine-grained message-driven execution. However, existing communication libraries, most notably MPI, were not originally designed to support massive thread-level concurrency or highly asynchronous control flows, making them a major bottleneck in scaling AMT systems on modern multicore and heterogeneous architectures. Although recent MPI extensions, such as Virtual Communication Interfaces (VCIs) and Continuations, attempt to address these limitations, they remain constrained by backward compatibility burdens and legacy runtime designs. This dissertation rethinks the design of communication libraries from the ground up, introducing the Lightweight Communication Interface (LCI), a communication library built from the ground up for asynchronous multithreaded communication. LCI provides a unified interface that supports common point-to-point primitives and diverse synchronization mechanisms, along with flexible controls for incrementally fine-tuning resource assignment and runtime behavior. It features a threading-efficient runtime built on atomic data structures, fine-grained non-blocking locks, and low-level network insights. It achieves scalable multithreaded performance without sacrificing simplicity or portability. We evaluate LCI and its integrations within two major AMT runtimes, including HPX and Charm++. In HPX, the LCI parcelport achieves higher message rates and lower latencies than its best MPI-based designs while requiring fewer communication contexts. In Charm++, we collaborated with the Charm++ development team to develop ReConverse, a modern reimplementation of the Converse communication substrate featuring a modular backend abstraction. The LCI backend within ReConverse eliminates the need for a dedicated communication thread and delivers order-of-magnitude performance improvements in latency, message rate, and real-world application throughput compared to its existing UCX and OFI machine layers. The results demonstrate that the fundamental barriers to multithreaded scalability stem not from hardware limitations but from legacy software abstractions. By relaxing unnecessary synchronization constraints and co-designing the interface and runtime around asynchrony, it is possible to achieve both flexibility and efficiency at scale. LCI and its integrations into HPX and Charm++ establish a practical path toward next-generation communication systems that treat multithreading and message-driven execution not as challenges, but as first principles for performance and programmability in high-performance computing.

Graduation Semester

2025-12

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/132527

Copyright and License Information

Copyright 2025 Jiakun Yan

Owning Collections