Workshops Archive

DECARD: A Runtime for Distributed and Heterogeneous Architectures Based on Dataflow


Workshop: 15th International Workshop on Runtime and Operating Systems for Supercomputers (ROSS 2026)

Authors: Diego Andres Roa Perdomo (University of Delaware, CAPSL) and Xiaoming Li (University of Delaware)

Abstract: Distributed heterogeneous architectures improve performance by mapping application components to specialized resources, but exposing architectural details increases programming and runtime complexity. Program Execution Models (PXMs) abstract application dependencies from scheduling and data-movement mechanisms. The Codelet Model is a dataflow-inspired PXM based on explicit dependencies among codelets grouped into Threaded Procedures.

This paper presents DECARD (Distributed Execution Codelet Accelerated Runtime for Dataflow), an extension of the Codelet Model for distributed heterogeneous systems. DECARD defines hierarchical components for node-level scheduling, communication, codelet scheduling, and execution, implemented using C++, OpenMP, and MPI. Evaluation shows a stable runtime cost of 1.14--1.23~$\mu$s per codelet, 10.55$\times$ shared-node speedup from 4 to 48 cores with 87.95% efficiency, and 96.67% distributed scaling efficiency across five 48-core nodes. Communication and 2D stencil experiments further characterize token transfers and application-level execution, establishing DECARD as a platform for studying distributed dataflow execution.


Back to 15th International Workshop on Runtime and Operating Systems for Supercomputers (ROSS 2026) Archive Listing

Back to Full Workshop Archive Listing