Instant scholar - Falcon: A graph manipulation language for distributed heterogeneous systemsPhD thesis by Unnikrishnan Cheramangalath

Falcon: Bridging Graph Analytics and Heterogeneous Computing In an era when data is expanding faster than ever, the ability to extract patterns and relationships from massive datasets defines the core of computational intelligence. Graphs — abstract representations of relationships between entities — are at the heart of this effort, underpinning everything from social network analysis to genomic research and web indexing. However, the challenge lies not only in modelling these graphs but in processing them efficiently across large, diverse computing infrastructures. In his doctoral thesis titled “Falcon: A Graph Manipulation Language for Distributed Heterogeneous Systems,” Unnikrishnan Cheramangalath presents an ambitious and technically sophisticated framework designed to make graph analytics more efficient, scalable, and accessible. Conducted under the Faculty of Engineering, this PhD work introduces Falcon, a domain-specific language (DSL) that reimagines how large-scale graph computations are expressed and executed across distributed, heterogeneous computing systems. Graph algorithms — such as PageRank, shortest path search, and connected components — are fundamental in analysing networks. Yet, executing these algorithms efficiently over heterogeneous systems (combinations of CPUs, GPUs, and distributed nodes) presents a formidable problem. Existing systems often force developers to grapple with low-level details: data partitioning, communication management, load balancing, and synchronisation. These complexities not only slow down development but also make it hard to scale applications efficiently. Cheramangalath’s thesis identifies this gap — the absence of a unified, high-level abstraction that can express graph algorithms once and execute them anywhere — as the motivation for Falcon. At its core, Falcon is a graph manipulation language — a domain-specific programming language designed for describing graph algorithms in a concise, expressive, and hardware-agnostic way. Falcon allows programmers to specify what to compute without explicitly describing how to compute it. The underlying compiler and runtime system handle the “how”: deciding how data is partitioned, how communication between nodes is managed, and how parallel execution is orchestrated across different hardware types. This separation of algorithm logic from system-level detail makes Falcon both developer-friendly and performance-oriented, bringing the benefits of abstraction without sacrificing speed. Cheramangalath’s work stands out for its focus on heterogeneity — environments where CPUs, GPUs, and distributed nodes coexist. Falcon’s compiler and runtime system are designed to target such architectures seamlessly. When a Falcon program is executed, the compiler automatically: This model enables users to write a single high-level program that can execute efficiently on a laptop, a multi-GPU workstation, or a large distributed cluster. Falcon’s architecture combines the conceptual clarity of a DSL with the execution efficiency of modern distributed systems. The system is built on three key pillars: This architecture gives Falcon flexibility without compromising the performance that high-end graph analytics demand. Falcon was tested across a range of standard graph algorithms — including Breadth-First Search (BFS), Single Source Shortest Path (SSSP), PageRank, and Connected Components — demonstrating robust performance and scalability. The results show that Falcon-generated code performs competitively against hand-optimised C++, CUDA, or MPI implementations. Moreover, because Falcon automates data distribution and parallel coordination, it reduces programming complexity by an order of magnitude — allowing researchers and developers to focus on algorithm design rather than system management. Cheramangalath’s research is a major contribution to the ongoing quest to make high-performance computing (HPC) accessible to data scientists and developers who are not experts in parallel or distributed programming. By blending domain-specific expressiveness with system-level intelligence, Falcon represents a step toward the next generation of graph analytics tools that can operate across diverse computing environments — from edge systems to cloud-scale clusters. Future work, as noted in the thesis, may extend Falcon to integrate with emerging architectures, support dynamic graphs (where nodes and edges evolve in real-time), and provide more advanced compiler optimisations for heterogeneous memory hierarchies. Unnikrishnan Cheramangalath’s doctoral thesis presents Falcon as a groundbreaking attempt to unify the fragmented landscape of graph computation. It stands as a testament to how language design, compiler engineering, and distributed systems can converge to solve one of the most pressing computational challenges of the data age: scalable, efficient, and portable graph analytics. By abstracting away the low-level complexities and automating heterogeneous execution, Falcon opens the door for researchers and engineers to think algorithmically — and let the system think architecturally.