Workshops Archive

Drava: An Event-Driven Runtime for Scientific Pipelines with Agentic Configuration Optimization


Workshop: XLOOP 2026: The 8th Annual Workshop on Extreme-Scale Experiment-in-the-Loop Computing

Authors: Ahmedur Rahman Shovon, Romain Pereira, Xingfu Wu, Brice Videau, and Swann Perarnau (Argonne National Laboratory)

Abstract: Across scientific domains, detectors increasingly produce data faster than it can be transmitted, forcing analysis to run near the instrument. The analysis uses multi-stage pipelines that may include reconstruction and inference under latency and throughput constraints. Current solutions leave key performance choices such as batch size, inter-stage data transport, and resource placement to manual tuning that is brittle and time consuming. We introduce Drava, an event-driven runtime for edge scientific streaming pipelines. Developers express stages as reusable callbacks, while Drava manages scheduling, batching, data movement, and transport-aware execution across heterogeneous resources. On ptychography and tomography pipelines, Drava sustains 31 kHz on the runtime path and runs up to 2.36× faster than a PvaPy baseline that drops frames beyond 2 kHz. Drava remains ~2.6× faster than hand-tuned loss-free PvaPy, while its observability enables agentic optimization that finds a configuration 2.24× faster than the best manual one and exposes energy–throughput trade-offs.


Back to XLOOP 2026: The 8th Annual Workshop on Extreme-Scale Experiment-in-the-Loop Computing Archive Listing

Back to Full Workshop Archive Listing